Low-cost air quality sensors

– are they reliable?

Alastair C Lewis, Peter M Edwards, James D Lee, Marvin Shaw,

Mat J Evans, Sarah J Moller, Katie Smith.

Wolfson Atmospheric Chemistry Laboratories, University of York, Heslington, York, UK.

Based on:

Lewis et al., Faraday Discussions. 189, 85-103, 2015

Lewis and Edwards, Nature, 535, 29-31, 2016

Sensors – a revolution in air pollution measurement?

o Current approach offers high quality measurements but poor spatial coverage.

o Distributed sensors could greatly improve coverage – personal exposure.

o Relies on assumption that the sensor data is fit for purpose.

Some of the hype......

A crowded marketplace

What is in the box?

SensorMicro-electro-mechanical (MEMS) device

Metal oxide

~ £5

~ 1960

Electrochemical

/ voltammetric

~ £50

~ 1980

Photochemical

~ £100

~ 1990

Micro-optical

> £100

~ 2000

“Hype Cycle” model used by

Gartner since 1995 and the

“Technology Adoption Lifecycle”

model popularized by Rogers

and Moore

Hype Cycle and Technology

Adoption Lifecycle

“Hype Cycle” model used by

Gartner since 1995 and the

“Technology Adoption Lifecycle”

model popularized by Rogers

and Moore

Hype Cycle and Technology

Adoption Lifecycle

Peer Review Standardisation

Universities

Industry

Universities

Institutes

National

Agencies

National

Regulators

Local

RegulatorsTraditional

Model

“Hype Cycle” model used by

Gartner since 1995 and the

“Technology Adoption Lifecycle”

model popularized by Rogers

and Moore

Hype Cycle and Technology

Adoption Lifecycle

Promotion Peer review

Start-ups Public Universities ?? ??Sensors

Model

Twenty sensor inter-comparison

o Reference methods used UV, Chemiluminescence, GC, TEOM-FDMS

o Devices initially calibrated to the reference value (e.g. slope applied on 11 Oct)

Ozone intercomparison – a success story?

Ozone sensors in more detail

o Collective accuracy is good, but individual accuracy is poor.

o Useable for research?? Probably.

o For the public?? They are not overtly misleading, since no collective bias

Highest sensor

Lowest sensor

Reference UV method

25th – 75 %-tile sensors

NO2– sensor-to-sensor variability

o Bias of 3.2 ± 1.7 – sensors over-

measure vs. reference.

o Poorer agreement on trends –some

other parameter e.g. CO2?

o Misleading public data – widespread

exceedances indicated

Highest sensor

Lowest sensor

Reference CLD method

25th – 75 %-tile sensors

Working electrode responses (in mV ppb-1 of co-

pollutant) induced by the presentation of co-

pollutants in zero air across five electrochemical

sensors

Potentially significant interference

Sensor interferences from co-pollutants

Co-pollutants

Six sensor test flow cell

NO2 sensor interference example

o NO2 electrochemical sensor has a small cross sensitivity to CO2

o But CO2 is generally in huge excess to NO2 in urban air

o At low [NO2] the electrochemical sensor is primarily sensing CO2

Sensor is measuring NO2

Sensor is measuring CO2

Sensor is measuring a

bit of both

Can we separate the signals?

o Interferences from other variables are the key sensor weakness

o These can interact with one another in non-linear ways

Constant 5 ppb isoprene,

variable RH

Temp

RH

Voltage

[VOC]

o Interferences from other variables are the key sensor weakness

o These can interact with one another in non-linear ways

Constant 5 ppb isoprene,

variable RH

Boosted Regression Tree

Original Signal

Gaussian Process Emulator

Separating signals from interferences

Not all sensors components are equal – e.g. PM

o Large observed variability in sensor performance.

o Not always obvious

which sensors /

technologies are used

in commercial units.

Sen

sor

Overcoming sensor variability – sensor clusters

o Even for the best sensors – e.g. O3, sensor-to-sensor reproducibility is poor.

o Highly reproducible sensors have the most uses even if the measurement is

operationally defined. E.g. for data assimilation.

o Sensor averages may be the way forward

Each sensor has a unique response

Sensor-to-sensor responsevaries by up to 200% at 2-sigma

8 x cluster ~ £80

Cluster unit response

varies by up to ~20% at

2-sigma.

Use dynamic

median

sensor value

Methodology:

Discard individual sensors

outside the interquartiles.

Use the resulting median

value from the remaining

sensors.

Sensor cluster reproducibility

A sensor or an instrument?

• There are alternative technological

pathways through lose device

miniturisation

• Often a more evolutionary approach

to measurements

• Use ‘traditional’ analytical

approaches, but smaller and cheaper

• Examples include GC or IMS – for

VOCs, optical for CO2, O3 or NO2

Conclusions

• Low cost sensors are an exciting opportunity.

• Wide range of sub-components but of variable quality.

• Publication bias, few independent tests reported, limited

academic publication.

• Cross-interferences from other pollutants.

• Unit – to – unit reproducibility can be very poor.

• Can generate misleading information - over-reporting is

commonplace.

• ‘Miniaturized’ instruments using known methodologies look

promising, e.g. OPCs.

• Statistical methods offer considerable promise, if backed up

by lab work.

• Clusters of sensors are more reproducible unit-to-unit

• But buyer beware!