Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
SlideShare a Scribd company logo

Low Cost Sensors to Measure Air Quality

•Download as PPTX, PDF•
•

This document summarizes the development and deployment of low-cost sensor networks to monitor urban air quality. It outlines the objective to better understand air pollution in megacities using dense sensor arrays. It then describes current research at MIT using nonparametric regression techniques and particle counter modeling to improve data from low-cost sensors. Finally, it provides examples of sensor networks deployed in Hawaii, Boston, and Delhi to measure pollutants like PM, O3, CO, and VOCs.

1 of 40

1

MONITORING URBAN AIR QUALITY USING DISTRIBUTED NETWORKS OF LOW-COST SENSORS DAVID H. HAGAN, JESSE H. KROLL SUMMER 2017

2

WHO AM I? …a design engineer and atmospheric chemist with experience across academia and industry.

3

OBJECTIVE ...to develop a more complete understanding of urban atmospheric chemistry in megacities using dense networks of low-cost atmospheric measurement devices. ...to inform results-driven public policy that will have lasting change.

4

OUTLINE What is air pollution and why do we care? Current research direction at MIT MIT AQ Monitoring Networks

5

Values from Hallquist et al, 2009 GASES / VAPORS The carbon life-cycle is complicated Particles, SO2, NOx, CO, VOCs EMISSIONS ( 1350 TgC yr -1 ) PARTICLES FORMED FROM GASES ( 60-240 TgC yr -1 )OXIDATION PARTICLE AND VAPOR DEPOSITION ( ~1000 TgC yr -1 ) OXIDATION TO CO2 ( 300 - 500 TgC yr -1 )

6

Large fraction of PM1 in Kanpur is secondary NR-PM1 Chemical Composition (AMS) AMS-PMF Factors for OA Kumar et al, 2016

7

Common/Regulated Pollutants POLLUTANT INDIA CPCB, 24H MEAN US EPA, 24H MEAN SO2 30.5 ppb 75 ppbc NO - - NO2 42.5 ppb 100 ppbc O3 50 ppba 70 ppba CO 1.75 ppma 9 ppm PM1 - - PM2.5 60 μgm-3 35 μgm-3 PM10 100 μgm-3 150 μgm-3 VOLATILE ORGANIC COMPOUNDS Benzene 1.57 ppbb - Ammonia 575 ppb - **There are other gas and particle phase pollutants that are not of interest to us at this time (metals, etc)** a 8-Hour Mean b Annual Mean c 1-Hour Mean

8

WHY DO WE CARE? AIR POLLUTION AND HUMAN HEALTH

9

Air pollution impedes quality of life 3.2M attributable deaths from ambient PM2.5 pollution - (GBD, 2010) adapted from Lim, et al. , Lancet (2012)

10

Air pollution decreases life expectancy • 3.2M premature mortalities world-wide1 • PM2.5 and O3 reduce life expectancy by 6.3 years in Delhi and 3.4 years across IN2 • $640B decrease in GDP (in IN) due to mortalities2 [1] Apte et al (2015) [2] Ghude et al

11

Particle diameter is important for health effects Nasal/Pharyngeal • peak at ~1 nm, 1 um Oberdorster, et al. (2005) Nasal/Pharyngeal • peak at ~5 nm Alveolar • peak at ~20 nm

12

Monitoring is sparse Area = 1484 km2 Pop. (2011) = 16.3M DPCC, 2015 Data from OpenAQ through DPCC COV MIN MAX SO2 13 % 0 ppb 115 ppb CO 65 % 0 ppm 33 ppm NO2 66 % 0 ppb 182 ppb O3 65 % 0 ppb 461 ppb PM2.5 65 % 1 μgm-3 985 μgm-3 DELHI DPCC RESULTS (2016)

13

Current efforts are insufficient Air Quality Egg Perkin Elmer ELM

14

“Data of poor or unknown quality is less useful than no data since it can lead to wrong decisions” Snyder, et al. “The Changing Paradigm of Air Pollution Monitoring”, ES&T 2013.

15

Low Cost Sensors to Measure Air Quality

16

Low Cost Sensors to Measure Air Quality

17

CURRENT RESEARCH AT MIT DESIGN AND DEPLOYMENT OF LOW-COST SENSOR NETWORKS

18

Research objectives include… 1) Can nonparametric regression techniques be used to “get- around” current modeling capabilities for electrochemical sensors? 2) How do perturbations in ambient aerosol distributions affect low-cost particle counters? 3) Can we use an array of broadband VOC sensors to obtain real-time information about the distribution of volatiles by chemical class? 4) Can we obtain information about particle formation and growth using a multi-pollutant air quality sensor?

19

NONPARAMETRIC REGRESSION AND ELECTROCHEMICAL SENSORS SO2 MEASUREMENTS IN HAWAII

20

State of electrochemical sensors MOTIVATION • current algorithms have proven ineffective • mostly LR • RF methods can’t extrapolate (Zimmerman et al) • Cross-sensitive species and environmental artifacts are very important • most methods are not real-time (Mead et al) • Long-term sensitivity decay is unknown APPROACH • Develop a method that combines the pros and cons of parametric (LR) and nonparametric (kNN) algorithms into one. • Validate using laboratory experiments and long-term co-location.

21

Temperature effects on EC sensors At low concentrations, T is very important Hagan et al, 2017

22

k-Nearest Neighbors Hybrid Regression • use LR at high concentrations • use kNN at low concentrations • combine using a linear combination of the two algorithms Hagan et al, 2017

23

We find no significant decay over 18 weeks Hagan et al, 2017

24

LOW-COST PARTICLE COUNTERS MODELING AND EXPERIMENTAL RESULTS

25

Optical particle counters have limitations MOTIVATION • widespread adoption in industry and through community-led initiatives • potential for rich dataset on PM/PN • optical methods know nothing about optical properties • most OPCs cannot see most particles APPROACH Model the response of OPCs to various aerosol distributions to probe the uncertaintly according to the underlying aerosol physics.

26

OPCs come in many shapes and sizes Sharp GP2Y1010AUOF Dmin : ~ 0.5 μm Dmax : unknown Nbins: 1 Cost: ~$10-20 Output: Analog signal ∝ Conc. Dylos DC1100 Pro Dmin: 0.5 μm Dmax: 10 μm Nbins: 2 Cost: ~$200 Output: #/ft3 Alphasense OPC-N2 Dmin: 0.38 μm Dmax: 17.5 μm Nbins: 16 Cost: ~$400 Output: #/cm3

27

REPRESENTATIVE AEROSOL DISTRIBUTION SIMULATED OPC • the sum of n lognormal modes, each described by: • Ni: total number of particles • Dpi: median diameter • σi: standard deviation • described by parameters: • Number of Bins • Minimum cutoff diameter (Dmin) • Maximum cutoff diameter (Dmax) • Counting efficiency f(RI, ρ, κ, composition, morphology)

28

Actual aerosol distribution | Cambridge, MA

29

Actual aerosol distribution | Cambridge, MA

30

Metrics of evaluation for simulated OPCs NUMBER-VOLUME CORRELATION • Assume number concentration is proportional to mass/volume (Wang et al (2015), DaCunto et al (2015)) • Ex. Sharp, Shinyei, Samyoung, etc • Limitations: Nm/Va changes as the underlying aerosol distribution changes INTEGRATED VOLUME • Sum the inferred volumes on a bin-by-bin basis • Ex. Alphasense OPC-N2, Dylos (sort of…) • Limitations: all volume below Dmin is lost

31

Nbins and Dmin are important Adding bins for a single distribution Changing distribution (GM) for an OPC Hagan et al, 2017 (in prep)

32

Initial conclusions and next steps • multi-bin resolution is needed to capture nuance of changing particle size distribution • sensitivity and accuracy of both methods improves with lower Dmin • lowering Dmin is more important than more bins (for design) Is there a way we can get around these limitations without needed an OPC that can see down to <100 nm?

33

Measuring PN in Delhi with a 658-nm OPC dV/dlogDp Dp um

34

MIT AQ MONITORING NETWORKS TO BOSTON, HAWAII, DELHI, & BEYOND

35

DEVICE PLATFORM Control (OTA Updates, etc) https://tatacenter-airquality.mit.edu  Database server is built with MySQL and WebScaleSQL  Backend is built with Flask using websockets  Everything flows through an NGINX reverse proxy load balancer  RESTful API for 3rd party integration/data retrieval

36

HI-VOG Network, Hawai’i USA • 12 SO2 sensors spread across elementary/middle schools and DOH sites • Launched (for good) Jan. 2017 • ~40 more sensors will be added in 2017-2018 along with CO and PM measurements • 6 sensors are permanently co-located with DOH reference stations for validation • Initial results are published in AMTD Hagan et al, 2017

37

Dorchester Air Quality Sensor Study, Boston USA • 5 multi-pollutant sensors spread across East Boston • 2 at MA DEP sites for co-location • Operational since 2015 • Measuring: • PM2.5, PM10 • O3, CO, NO, NO2 • In collaboration with Aerodyne Research and the Harvard School of Public Health Cross et al, 2017

38

Multi-Pollutant Air Quality Sensor (MPAQS), Delhi IN • 24 sensors (12 co-located, 12 at private company) • Engineered specifically for Delhi (high particle loadings in winter) • Running since 2015 (multiple versions) • co-located instruments include E-BAM (PM2.5), ACSM (PM1 composition), Aetholometer (BC), SMPS (size distribution), 2BTech 202 (O3) • work with Josh Apte (UT Austin)

39

Low Cost Sensors to Measure Air Quality

40

@dhhagan @dhagan7 dhagan@mit.edu

More Related Content

Low Cost Sensors to Measure Air Quality

  • 1. MONITORING URBAN AIR QUALITY USING DISTRIBUTED NETWORKS OF LOW-COST SENSORS DAVID H. HAGAN, JESSE H. KROLL SUMMER 2017
  • 2. WHO AM I? …a design engineer and atmospheric chemist with experience across academia and industry.
  • 3. OBJECTIVE ...to develop a more complete understanding of urban atmospheric chemistry in megacities using dense networks of low-cost atmospheric measurement devices. ...to inform results-driven public policy that will have lasting change.
  • 4. OUTLINE What is air pollution and why do we care? Current research direction at MIT MIT AQ Monitoring Networks
  • 5. Values from Hallquist et al, 2009 GASES / VAPORS The carbon life-cycle is complicated Particles, SO2, NOx, CO, VOCs EMISSIONS ( 1350 TgC yr -1 ) PARTICLES FORMED FROM GASES ( 60-240 TgC yr -1 )OXIDATION PARTICLE AND VAPOR DEPOSITION ( ~1000 TgC yr -1 ) OXIDATION TO CO2 ( 300 - 500 TgC yr -1 )
  • 6. Large fraction of PM1 in Kanpur is secondary NR-PM1 Chemical Composition (AMS) AMS-PMF Factors for OA Kumar et al, 2016
  • 7. Common/Regulated Pollutants POLLUTANT INDIA CPCB, 24H MEAN US EPA, 24H MEAN SO2 30.5 ppb 75 ppbc NO - - NO2 42.5 ppb 100 ppbc O3 50 ppba 70 ppba CO 1.75 ppma 9 ppm PM1 - - PM2.5 60 μgm-3 35 μgm-3 PM10 100 μgm-3 150 μgm-3 VOLATILE ORGANIC COMPOUNDS Benzene 1.57 ppbb - Ammonia 575 ppb - **There are other gas and particle phase pollutants that are not of interest to us at this time (metals, etc)** a 8-Hour Mean b Annual Mean c 1-Hour Mean
  • 8. WHY DO WE CARE? AIR POLLUTION AND HUMAN HEALTH
  • 9. Air pollution impedes quality of life 3.2M attributable deaths from ambient PM2.5 pollution - (GBD, 2010) adapted from Lim, et al. , Lancet (2012)
  • 10. Air pollution decreases life expectancy • 3.2M premature mortalities world-wide1 • PM2.5 and O3 reduce life expectancy by 6.3 years in Delhi and 3.4 years across IN2 • $640B decrease in GDP (in IN) due to mortalities2 [1] Apte et al (2015) [2] Ghude et al
  • 11. Particle diameter is important for health effects Nasal/Pharyngeal • peak at ~1 nm, 1 um Oberdorster, et al. (2005) Nasal/Pharyngeal • peak at ~5 nm Alveolar • peak at ~20 nm
  • 12. Monitoring is sparse Area = 1484 km2 Pop. (2011) = 16.3M DPCC, 2015 Data from OpenAQ through DPCC COV MIN MAX SO2 13 % 0 ppb 115 ppb CO 65 % 0 ppm 33 ppm NO2 66 % 0 ppb 182 ppb O3 65 % 0 ppb 461 ppb PM2.5 65 % 1 μgm-3 985 μgm-3 DELHI DPCC RESULTS (2016)
  • 13. Current efforts are insufficient Air Quality Egg Perkin Elmer ELM
  • 14. “Data of poor or unknown quality is less useful than no data since it can lead to wrong decisions” Snyder, et al. “The Changing Paradigm of Air Pollution Monitoring”, ES&T 2013.
  • 17. CURRENT RESEARCH AT MIT DESIGN AND DEPLOYMENT OF LOW-COST SENSOR NETWORKS
  • 18. Research objectives include… 1) Can nonparametric regression techniques be used to “get- around” current modeling capabilities for electrochemical sensors? 2) How do perturbations in ambient aerosol distributions affect low-cost particle counters? 3) Can we use an array of broadband VOC sensors to obtain real-time information about the distribution of volatiles by chemical class? 4) Can we obtain information about particle formation and growth using a multi-pollutant air quality sensor?
  • 20. State of electrochemical sensors MOTIVATION • current algorithms have proven ineffective • mostly LR • RF methods can’t extrapolate (Zimmerman et al) • Cross-sensitive species and environmental artifacts are very important • most methods are not real-time (Mead et al) • Long-term sensitivity decay is unknown APPROACH • Develop a method that combines the pros and cons of parametric (LR) and nonparametric (kNN) algorithms into one. • Validate using laboratory experiments and long-term co-location.
  • 21. Temperature effects on EC sensors At low concentrations, T is very important Hagan et al, 2017
  • 22. k-Nearest Neighbors Hybrid Regression • use LR at high concentrations • use kNN at low concentrations • combine using a linear combination of the two algorithms Hagan et al, 2017
  • 23. We find no significant decay over 18 weeks Hagan et al, 2017
  • 24. LOW-COST PARTICLE COUNTERS MODELING AND EXPERIMENTAL RESULTS
  • 25. Optical particle counters have limitations MOTIVATION • widespread adoption in industry and through community-led initiatives • potential for rich dataset on PM/PN • optical methods know nothing about optical properties • most OPCs cannot see most particles APPROACH Model the response of OPCs to various aerosol distributions to probe the uncertaintly according to the underlying aerosol physics.
  • 26. OPCs come in many shapes and sizes Sharp GP2Y1010AUOF Dmin : ~ 0.5 μm Dmax : unknown Nbins: 1 Cost: ~$10-20 Output: Analog signal ∝ Conc. Dylos DC1100 Pro Dmin: 0.5 μm Dmax: 10 μm Nbins: 2 Cost: ~$200 Output: #/ft3 Alphasense OPC-N2 Dmin: 0.38 μm Dmax: 17.5 μm Nbins: 16 Cost: ~$400 Output: #/cm3
  • 27. REPRESENTATIVE AEROSOL DISTRIBUTION SIMULATED OPC • the sum of n lognormal modes, each described by: • Ni: total number of particles • Dpi: median diameter • σi: standard deviation • described by parameters: • Number of Bins • Minimum cutoff diameter (Dmin) • Maximum cutoff diameter (Dmax) • Counting efficiency f(RI, ρ, κ, composition, morphology)
  • 28. Actual aerosol distribution | Cambridge, MA
  • 29. Actual aerosol distribution | Cambridge, MA
  • 30. Metrics of evaluation for simulated OPCs NUMBER-VOLUME CORRELATION • Assume number concentration is proportional to mass/volume (Wang et al (2015), DaCunto et al (2015)) • Ex. Sharp, Shinyei, Samyoung, etc • Limitations: Nm/Va changes as the underlying aerosol distribution changes INTEGRATED VOLUME • Sum the inferred volumes on a bin-by-bin basis • Ex. Alphasense OPC-N2, Dylos (sort of…) • Limitations: all volume below Dmin is lost
  • 31. Nbins and Dmin are important Adding bins for a single distribution Changing distribution (GM) for an OPC Hagan et al, 2017 (in prep)
  • 32. Initial conclusions and next steps • multi-bin resolution is needed to capture nuance of changing particle size distribution • sensitivity and accuracy of both methods improves with lower Dmin • lowering Dmin is more important than more bins (for design) Is there a way we can get around these limitations without needed an OPC that can see down to <100 nm?
  • 33. Measuring PN in Delhi with a 658-nm OPC dV/dlogDp Dp um
  • 34. MIT AQ MONITORING NETWORKS TO BOSTON, HAWAII, DELHI, & BEYOND
  • 35. DEVICE PLATFORM Control (OTA Updates, etc) https://tatacenter-airquality.mit.edu  Database server is built with MySQL and WebScaleSQL  Backend is built with Flask using websockets  Everything flows through an NGINX reverse proxy load balancer  RESTful API for 3rd party integration/data retrieval
  • 36. HI-VOG Network, Hawai’i USA • 12 SO2 sensors spread across elementary/middle schools and DOH sites • Launched (for good) Jan. 2017 • ~40 more sensors will be added in 2017-2018 along with CO and PM measurements • 6 sensors are permanently co-located with DOH reference stations for validation • Initial results are published in AMTD Hagan et al, 2017
  • 37. Dorchester Air Quality Sensor Study, Boston USA • 5 multi-pollutant sensors spread across East Boston • 2 at MA DEP sites for co-location • Operational since 2015 • Measuring: • PM2.5, PM10 • O3, CO, NO, NO2 • In collaboration with Aerodyne Research and the Harvard School of Public Health Cross et al, 2017
  • 38. Multi-Pollutant Air Quality Sensor (MPAQS), Delhi IN • 24 sensors (12 co-located, 12 at private company) • Engineered specifically for Delhi (high particle loadings in winter) • Running since 2015 (multiple versions) • co-located instruments include E-BAM (PM2.5), ACSM (PM1 composition), Aetholometer (BC), SMPS (size distribution), 2BTech 202 (O3) • work with Josh Apte (UT Austin)

Editor's Notes

  1. CO: In India, the measurement is an annual mean (min. 104 measurements in a year) whereas in the US, it cannot be exceeded once!
  2. fig of where particles end up!
  3. Modeling, Characterization, and Deployment of MPAQS and VOC Instruments
  4. oveview of OPCs
  5. GIF of PN distribution
  6. How do they work (2 algos)
  7. Model results
  8. new algo approach
  9. introduce the MPAQS
  10. network structure
  11. network structure
  12. Laboratory Setup
  13. add linkedin