Content uploaded by Achim Dittler
Author content
All content in this area was uploaded by Achim Dittler on Dec 29, 2018
Content may be subject to copyright.
3
rd
International Conference on Atmospheric Dust - DUST2018
ProScience 5 (2018) 6-12
Available at www.scientevents.com/proscience/
Conference Logo – Information
Do not modify this area
ISSN: 2283-5954© 2018 The Authors. Published by Digilabs
Selection and peer-review under responsibility of DUST2018 Scientific Committee
DOI: 10.14644/dust.2018.002
Potential and Limitations of the Low-Cost SDS011
Particle Sensor for Monitoring Urban Air Quality
Matthias Budde1*, Almuth D. Schwarz2, Thomas Müller3, Bernd Laquai4,
Norbert Streibl5, Gregor Schindler1, Marcel Köpke1, Till Riedel1,
Achim Dittler2, Michael Beigl1
1Karlsruhe Institute of Technology (KIT), Institute of Telematics, Chair for Pervasive
Computing Systems / TECO, Karlsruhe, Germany
2Karlsruhe Institute of Technology (KIT), Institute for Mechanical Process Engineering and
Mechanics, Gas Particle Systems , Karlsruhe, Germany
3Leibnitz Institute for Tropospheric Research (TROPOS), Leipzig, Germany
4University of Stuttgart, Institute of Combustion and Power Plant Technology (IFK),
Department Air Quality Control, Stuttgart, Germany
5Independant Researcher, Germany
*budde@teco.edu
Abstract
In Particulate Matter (PM) monitoring, current laser scattering low-cost sensor generations exhibit
better stability than early sensor generations and feature internal digital processing to achieve more
accurate results. As a representative of this class of sensors, we examine the popular SDS011 PM
sensor. Previous work about co-location measurements between SDS011 sensors indicates that the
sensor delivers adequate correlation under “typical” conditions, but performs less well under other
ambient conditions, especially high humidity. To further explore the sensor's data quality in-depth, we
conducted a series of controlled experiments with high precision reference devices and different
aerosols, both polydisperse (ambient air, ammonium sulfate, soot) and monodisperse (polystyrene
particles of different sizes). We also present results from a longer-term comparison (days) of multiple
sensors, as well as the key influencing factors on uncertainty and assess the sensor’s potential and
limitations. Our findings show that a single sensor generally does not capture PM10 satisfactorily and
we discuss under which conditions PM2.5 readings reflect the ambient air quality adequately.
Keywords: Low-Cost Sensors, PM, Environmental Sensing, Air Quality, Aerosol Monitoring, Evaluation
1. Introduction
In Particulate Matter (PM) monitoring, developments towards incorporating distributed
sensing approaches using low-cost sensors are being made (Snyder et al., 2013). Research
on early generations of low-cost particle sensors compares them with official measurement
stations, showing that they can in principle capture the dynamics of ambient PM levels
(Budde et al., 2013; Holstius et al., 2014), but may suffer from low calibration stability, are
unable to differentiate size classes, and may be susceptible to other sources of error (Budde
et al., 2015). Current low-cost laser scattering sensors claim to exhibit better stability and
more accurate readings. While they are mostly designated as PM2.5 sensors, some also
output PM10 and/or PM1 values. As a representative of this class of sensors, we examine the
SDS011 sensor (Nova Fitness, 2015). It is already widely being used in deployments
around the world, ranging from sensor networks to grassroots citizen science projects, in
which volunteers have deployed hundreds of these sensors in urban areas.
In related work, co-location measurements between the SDS011 have been performed
(LUBW, 2017), indicating that the sensor signal exhibits adequate correlation under some
conditions (relative humidity of 20-50 % and PM10 mass concentrations < 20 μg/m³) but
performs less well under others, especially high humidity. To further explore the sensor's
quality in-depth, we present the key influencing factors on measurement uncertainty, along
with a series of experiments to appropriately assess its potential and limitations.
2. Experiments
We conducted lab experiments exposing the sensor(s) to increasing concentration levels
of monodisperse and polydisperse particles, both artificially created and using ambient air.
2.1 Monodisperse Particulates
In order to assess the influence of concentration and particle size on the sensor readings,
a series of experiments was carried out measuring inert monodisperse polystyrene particles
with SDS011 sensors in a lab setting, with a Palas PROMO 2000 device with welas 2100
sensor as reference. An aerosol generator made specifically for fine control of the
conditions under which particles are dispersed was used in all experiments (see Fig. 1,
right).
Fig. 1. Experimental setup of our tests with monodisperse polystyrene particles
DOI:10.14644/dust.2018.002
Conference Logo – Information
Do not modify this area
ISSN: 2283-5954© 2018 The Authors. Published by Digilabs
Selection and peer-review under responsibility of DUST2018 Scientific Committee
DOI: 10.14644/dust.2018.002
Potential and Limitations of the Low-Cost SDS011
Particle Sensor for Monitoring Urban Air Quality
Matthias Budde1*, Almuth D. Schwarz2, Thomas Müller3, Bernd Laquai4,
Norbert Streibl5, Gregor Schindler1, Marcel Köpke1, Till Riedel1,
Achim Dittler2, Michael Beigl1
1Karlsruhe Institute of Technology (KIT), Institute of Telematics, Chair for Pervasive
Computing Systems / TECO, Karlsruhe, Germany
2Karlsruhe Institute of Technology (KIT), Institute for Mechanical Process Engineering and
Mechanics, Gas Particle Systems , Karlsruhe, Germany
3Leibnitz Institute for Tropospheric Research (TROPOS), Leipzig, Germany
4University of Stuttgart, Institute of Combustion and Power Plant Technology (IFK),
Department Air Quality Control, Stuttgart, Germany
5Independant Researcher, Germany
*budde@teco.edu
Abstract
In Particulate Matter (PM) monitoring, current laser scattering low-cost sensor generations exhibit
better stability than early sensor generations and feature internal digital processing to achieve more
accurate results. As a representative of this class of sensors, we examine the popular SDS011 PM
sensor. Previous work about co-location measurements between SDS011 sensors indicates that the
sensor delivers adequate correlation under “typical” conditions, but performs less well under other
ambient conditions, especially high humidity. To further explore the sensor's data quality in-depth, we
conducted a series of controlled experiments with high precision reference devices and different
aerosols, both polydisperse (ambient air, ammonium sulfate, soot) and monodisperse (polystyrene
particles of different sizes). We also present results from a longer-term comparison (days) of multiple
sensors, as well as the key influencing factors on uncertainty and assess the sensor’s potential and
limitations. Our findings show that a single sensor generally does not capture PM10 satisfactorily and
we discuss under which conditions PM2.5 readings reflect the ambient air quality adequately.
Keywords: Low-Cost Sensors, PM, Environmental Sensing, Air Quality, Aerosol Monitoring, Evaluation
1. Introduction
In Particulate Matter (PM) monitoring, developments towards incorporating distributed
sensing approaches using low-cost sensors are being made (Snyder et al., 2013). Research
on early generations of low-cost particle sensors compares them with official measurement
stations, showing that they can in principle capture the dynamics of ambient PM levels
(Budde et al., 2013; Holstius et al., 2014), but may suffer from low calibration stability, are
unable to differentiate size classes, and may be susceptible to other sources of error (Budde
et al., 2015). Current low-cost laser scattering sensors claim to exhibit better stability and
more accurate readings. While they are mostly designated as PM2.5 sensors, some also
output PM10 and/or PM1 values. As a representative of this class of sensors, we examine the
SDS011 sensor (Nova Fitness, 2015). It is already widely being used in deployments
around the world, ranging from sensor networks to grassroots citizen science projects, in
which volunteers have deployed hundreds of these sensors in urban areas.
In related work, co-location measurements between the SDS011 have been performed
(LUBW, 2017), indicating that the sensor signal exhibits adequate correlation under some
conditions (relative humidity of 20-50 % and PM10 mass concentrations < 20 μg/m³) but
performs less well under others, especially high humidity. To further explore the sensor's
quality in-depth, we present the key influencing factors on measurement uncertainty, along
with a series of experiments to appropriately assess its potential and limitations.
2. Experiments
We conducted lab experiments exposing the sensor(s) to increasing concentration levels
of monodisperse and polydisperse particles, both artificially created and using ambient air.
2.1 Monodisperse Particulates
In order to assess the influence of concentration and particle size on the sensor readings,
a series of experiments was carried out measuring inert monodisperse polystyrene particles
with SDS011 sensors in a lab setting, with a Palas PROMO 2000 device with welas 2100
sensor as reference. An aerosol generator made specifically for fine control of the
conditions under which particles are dispersed was used in all experiments (see Fig. 1,
right).
Fig. 1. Experimental setup of our tests with monodisperse polystyrene particles
ProScience 5 (2018) 6-12 7
Homogeneous watery suspensions containing the target particles are dispersed through a
venturi nozzle. The small droplets containing single particles then enter a chamber, where
they are mixed with a heated stream of dry, particle-free air, which dries the droplets and
dictates the temperature during the experiment. The air stream entering the suspension is
used to control the exiting concentration of the generator. The temperature and flow rate of
the hot air stream are used to control temperature and humidity during the experiment. The
stream exiting the generator contains the dispersed monodisperse particles that were
originally in the suspension. This stream passes through a diffusion dryer before entering
the measuring chamber, which was left open at the end, allowing to place SDS011 sensors
horizontally (with the fan facing towards the ceiling) and the reference sensor in the
chamber on the same height (Fig. 1, left). The particles used in these experiments were
inert, non-hygroscopic, monodisperse polystyrene particles with a density of 1.055 g/cm³.
To investigate the influence of particle size and concentration, monodisperse particles of
different sizes (0.3, 1, 2, 5 and 10 µm respectively) were used in separate experiments.
Various concentrations of up to approximately 2500 µg/m³ were used for each particle size.
The experiments were conducted at ambient pressure and a relative humidity of
approximately 45 %. Temperatures during the experiment were kept between 20 and 22 °C.
All sensors measured the concentration constantly for at least 20 minutes each to ensure
stable concentration conditions. The data presented in this paper shows the results of one
SDS011 sensor only. Other SDS011 sensors (four tested in total) showed the same trends.
Fig. 2. SDS011 readings vs. reference for different monodisperse particle sizes for PM2. 5 (left) and PM10 (right)
PM2.5 and PM10 readings of the SDS011 sensor deviate from the reference, depending
on particle size and concentration (see Fig. 2). For 0.3 µm particles, PM2.5 readings (Fig. 2,
left) are similar to those of the reference up to a reference concentration of 300 μg/m³. For
concentrations above this value, the SDS011 reports lower values for 0.3 µm than the welas
2100. For 1 μm particles, too high PM2.5 values are reported by the SDS011 for PM2.5
concentrations between 300 µg/m³ and 800 µg/m³. Below they fit well to the reference,
above PM2.5 concentrations are underestimated. 2 µm particles are strongly underestimated
by the SDS011 in the PM2.5 value. The PM2.5 results for 5 μm and 10 μm particles must be
treated with caution. For them, the reference device reported impurities in the size channels
between 0.2 µm and 0.5 µm, which lead to PM2.5 values below 10 µg/m³. The SDS011
shows much higher PM2.5 values (up to ~100 µg/m³), but it remains unclear if it reacts
stronger to the impurities or if it wrongly includes the 5 µm and 10 µm particles.
In the SDS011’s PM10 readings (Fig. 2, right), particles of 5 μm and 10 μm are severely
underestimated. PM10 readings for 0.3 µm particles are similar for both the SDS011 and the
reference. 1 µm particles are overestimated in the SDS011 PM10 value above concentrations
of 300 µg/m³. When we look at the PM10 values for 2 µm particles, we can see a peculiar
behavior. As we presented before, the SDS011 sensor underestimates the mass for 2 μm
particles in its PM2.5 values, indicating that it either does not follow the same PM2.5 curve as
our reference or assumes a too low density for the particles. However, for PM10, the
SDS011 and the welas 2100 report similar values for 2 μm particles.
Overall, the SDS011 shows good results for particles smaller than 5 µm for the PM10
value, as long as these are measured under dry, stable conditions, but does not sufficiently
consider particles with the sizes 5 µm and 10 µm. The sensor appears to deliver good PM2.5
readings for particle sizes 1 µm and 0.3 µm and good PM10 readings for particle sizes
0.3 µm and 2 µm for the material used in these experiments. It remains unclear how the
sensor’s internal processing estimates the particle sizes.
2.2 Polydisperse Particulates
We compared several SDS011 sensors with reference measurements in a lab setting at
the World Calibration Center for Aerosol Physics (WCCAP) at the Leibnitz Institute for
Tropospheric Research (TROPOS) that is operated in cooperation with the German Federal
Environmental Agency (UBA) and the World Meteorological Organization (WMO). We
exposed a set of 17 SDS011 sensors to different aerosols (ammonium sulfate, ambient air,
soot from a Jing Ltd. miniCAST propane gas burner, and zero air) in an airtight aluminum
chamber, depicted in Figure 3. The air outlet was connected to two reference devices: An
aerodynamic particle sizer (TSI APS model 3321) and a Scanning Mobility Particle Sizer
(SMPS) custom made at the WCCAP (Wiedensohler et al., 2012). This combination
enabled the measurement of 92 aerodynamic size channels between 10 nm and 20 µm.
Time resolution was one reading every 4.5 minutes. The SMPS samples the different
channels in a time multiplex fashion, i.e. one after another over a time of 4.5 minutes. This
is how long it takes to scan through all size channels. From these individual size channels,
we calculated the three size classes PM10, PM2.5 and PM1 as reference. Aerodynamic
particle diameters were converted to a geometric diameter assuming spherical particles and
using a particles density of 1.7 g/cm³.
Fig. 3. SDS011 sensors were placed inside an airtight aluminum measurement chamber (left), into which a varying
concentration of polydisperse particles was injected. An SMPS and a TSI APS were used as referencev
8ProScience 5 (2018) 6-12
Homogeneous watery suspensions containing the target particles are dispersed through a
venturi nozzle. The small droplets containing single particles then enter a chamber, where
they are mixed with a heated stream of dry, particle-free air, which dries the droplets and
dictates the temperature during the experiment. The air stream entering the suspension is
used to control the exiting concentration of the generator. The temperature and flow rate of
the hot air stream are used to control temperature and humidity during the experiment. The
stream exiting the generator contains the dispersed monodisperse particles that were
originally in the suspension. This stream passes through a diffusion dryer before entering
the measuring chamber, which was left open at the end, allowing to place SDS011 sensors
horizontally (with the fan facing towards the ceiling) and the reference sensor in the
chamber on the same height (Fig. 1, left). The particles used in these experiments were
inert, non-hygroscopic, monodisperse polystyrene particles with a density of 1.055 g/cm³.
To investigate the influence of particle size and concentration, monodisperse particles of
different sizes (0.3, 1, 2, 5 and 10 µm respectively) were used in separate experiments.
Various concentrations of up to approximately 2500 µg/m³ were used for each particle size.
The experiments were conducted at ambient pressure and a relative humidity of
approximately 45 %. Temperatures during the experiment were kept between 20 and 22 °C.
All sensors measured the concentration constantly for at least 20 minutes each to ensure
stable concentration conditions. The data presented in this paper shows the results of one
SDS011 sensor only. Other SDS011 sensors (four tested in total) showed the same trends.
Fig. 2. SDS011 readings vs. reference for different monodisperse particle sizes for PM2. 5 (left) and PM10 (right)
PM2.5 and PM10 readings of the SDS011 sensor deviate from the reference, depending
on particle size and concentration (see Fig. 2). For 0.3 µm particles, PM2.5 readings (Fig. 2,
left) are similar to those of the reference up to a reference concentration of 300 μg/m³. For
concentrations above this value, the SDS011 reports lower values for 0.3 µm than the welas
2100. For 1 μm particles, too high PM2.5 values are reported by the SDS011 for PM2.5
concentrations between 300 µg/m³ and 800 µg/m³. Below they fit well to the reference,
above PM2.5 concentrations are underestimated. 2 µm particles are strongly underestimated
by the SDS011 in the PM2.5 value. The PM2.5 results for 5 μm and 10 μm particles must be
treated with caution. For them, the reference device reported impurities in the size channels
between 0.2 µm and 0.5 µm, which lead to PM2.5 values below 10 µg/m³. The SDS011
shows much higher PM2.5 values (up to ~100 µg/m³), but it remains unclear if it reacts
stronger to the impurities or if it wrongly includes the 5 µm and 10 µm particles.
In the SDS011’s PM10 readings (Fig. 2, right), particles of 5 μm and 10 μm are severely
underestimated. PM10 readings for 0.3 µm particles are similar for both the SDS011 and the
reference. 1 µm particles are overestimated in the SDS011 PM10 value above concentrations
of 300 µg/m³. When we look at the PM10 values for 2 µm particles, we can see a peculiar
behavior. As we presented before, the SDS011 sensor underestimates the mass for 2 μm
particles in its PM2.5 values, indicating that it either does not follow the same PM2.5 curve as
our reference or assumes a too low density for the particles. However, for PM10, the
SDS011 and the welas 2100 report similar values for 2 μm particles.
Overall, the SDS011 shows good results for particles smaller than 5 µm for the PM10
value, as long as these are measured under dry, stable conditions, but does not sufficiently
consider particles with the sizes 5 µm and 10 µm. The sensor appears to deliver good PM2.5
readings for particle sizes 1 µm and 0.3 µm and good PM10 readings for particle sizes
0.3 µm and 2 µm for the material used in these experiments. It remains unclear how the
sensor’s internal processing estimates the particle sizes.
2.2 Polydisperse Particulates
We compared several SDS011 sensors with reference measurements in a lab setting at
the World Calibration Center for Aerosol Physics (WCCAP) at the Leibnitz Institute for
Tropospheric Research (TROPOS) that is operated in cooperation with the German Federal
Environmental Agency (UBA) and the World Meteorological Organization (WMO). We
exposed a set of 17 SDS011 sensors to different aerosols (ammonium sulfate, ambient air,
soot from a Jing Ltd. miniCAST propane gas burner, and zero air) in an airtight aluminum
chamber, depicted in Figure 3. The air outlet was connected to two reference devices: An
aerodynamic particle sizer (TSI APS model 3321) and a Scanning Mobility Particle Sizer
(SMPS) custom made at the WCCAP (Wiedensohler et al., 2012). This combination
enabled the measurement of 92 aerodynamic size channels between 10 nm and 20 µm.
Time resolution was one reading every 4.5 minutes. The SMPS samples the different
channels in a time multiplex fashion, i.e. one after another over a time of 4.5 minutes. This
is how long it takes to scan through all size channels. From these individual size channels,
we calculated the three size classes PM10, PM2.5 and PM1 as reference. Aerodynamic
particle diameters were converted to a geometric diameter assuming spherical particles and
using a particles density of 1.7 g/cm³.
Fig. 3. SDS011 sensors were placed inside an airtight aluminum measurement chamber (left), into which a varying
concentration of polydisperse particles was injected. An SMPS and a TSI APS were used as referencev
ProScience 5 (2018) 6-12 9
Figure 4 shows the time series of our experiments. Three individual sensors were
disregarded, as was the complete ambient air 2 session, due to technical errors. We can see
that the SDS011 generally captures dynamics well, but also that the readings have an offset
to the reference and that this offset differs for the individual sensors. This variance can also
be seen in the scatterplot in Figure 5.
Fig. 4. Time series of experiments with different polydisperse aerosols (left) and zoomed in for PM2.5 mass for the
first session of ambient air measurements (right). The reference is the solid black line
Fig. 5. Scatterplots: mean PM2.5 mass against the reference (left) and for all individual 14 valid sensors for the test
aerosol ambient air (right).
We see that the sensor exhibits overall good linearity, but also systematic misestimation
of the true concentration. On average, the SDS011 showed only 66 % of the PM2.5 in
ambient air. Individually, the readings ranged from 45 % to 85 %. For ammonium sulfate,
the SDS011s on average show only 48 % of the reference’s PM2.5. For pure soot, no usable
readings could be obtained, as was expected due strong absorption and low light scattering.
This is a general issue of optical measurements and not specific to the SDS011. It should be
mentioned that soot in ambient air is coated with other materials, e.g. organics and sulfates,
and scatters light. A general conversion factor how much soot is seen by the SDS011
cannot be given. For zero air, the sensors showed minimum PM2.5 values of ~0.3 µg/m³,
which seems acceptable for ambient air measurements. All results are presented in Table 1.
2.3 Other Factors
The other most noteworthy aspect that affects the SDS011 sensor is humidity.
Generally, the sensors are only specified to operate up to a relative humidity of 70% (Nova
Fitness, 2015). Above these values, large deviations must be expected. Figure 6a shows one
Table 1. Measurement results for the individual sensors (PM2.5 / PM10).
Sensor
Ambient air
Ammonium sulfate
Soot
Zero air
slope
intercept
slope
intercept
slope
intercept
baseline
stdev
1188476
0.85 / 0.85
-0.88 / -0.74
0.63 / 0.67
0.01 / 0.26
0 / 0
1.55 / 1.84
0.247 / 0.454
0.0020 / 0.0018
1190049
0.67 / 0.66
-0.66 / -0.50
0.50 / 0.53
0.19 / 0.40
0 / 0
1.21 / 1.39
0.303 / 0.400
0.0009 / 0.0002
1197947
0.61 / 0.60
-0.68 / -0.59
0.46 / 0.49
0.23 / 0.63
0 / 0
1.09 / 1.37
0.200 / 0.300
0.0007 / 0.0004
1200580
0.84 / 0.84
-0.60 / -0.59
0.62 / 0.66
0.24 / 0.41
-0.01 / -0.01
1.95 / 2.09
0.415 / 0.413
0.0017 / 0.0024
132097
0.65 / 0.64
-0.79 / -0.57
0.47 / 0.50
0.17 / 0.72
0 / 0
1.11 / 1.38
0.200 / 0.402
0.0003 / 0.0024
1547350
0.60 / 0.59
-0.80 / -0.76
0.47 / 0.51
-0.06 / 0.55
0 / 0
1.09 / 1.24
0.301 / 0.400
0.0004 / 0.0002
1547430
0.69 / 0.65
-0.98 / -0.67
0.48 / 0.53
0.03 / 0.94
0 / 0
1.26 / 1.68
0.300 / 0.800
0.0000 / 0.0003
1547436
0.64 / 0.60
-0.80 / -0.44
0.51 / 0.56
-0.12 / 0.34
0 / 0
1.27 / 1.70
0.393 / 0.700
0.0014 / 0.0000
1548172
0.48 / 0.45
-0.53 / 0.12
0.42 / 0.45
0.04 / 1.44
0 / 0
0.97 / 1.82
0.300 / 0.600
0.0002 / 0.0005
1548187
0.65 / 0.71
-0.52 / -0.48
0.47 / 0.53
0.36 / 1.08
0 / 0
1.28 / 1.84
0.407 / 0.601
0.0010 / 0.0008
1548303
0.66 / 0.69
-0.59 / -0.69
0.50 / 0.55
0.11 / 0.55
0 / 0
1.26 / 1.49
0.391 / 0.500
0.0012 / 0.0002
1581752
0.64 / 0.62
-0.62 / -0.41
0.396 / 0.699
0.0017 / 0.0013
158693
0.63 / 0.62
-0.76 / -0.62
0.46 / 0.49
0.21 / 0.53
0 / 0
1.08 / 1.37
0.200 / 0.401
0.0002 / 0.0004
1590613
0.52 / 0.52
-0.73 / -0.52
0.41 / 0.45
-0.13 / 0.65
0 / 0
0.94 / 1.15
0.300 / 0.400
0.0000 / 0.0004
Average
0.65 / 0.65
-0.71 / -0.53
0.49 / 0.53
0.10 / 0.65
day of data collected in-the-wild on a humid day. The graph shows that the sensor
drastically overestimates the particle concentration in the morning and evening hours when
humidity is high. There are two sources of error caused by humidity: Hygroscopic growth
of saline particles and condensed fog droplets of similar size as PM-relevant particles. We
observed that the overestimation is much stronger in fog, and less extreme in otherwise
high humidity or rain. These findings are in line with results published in parallel with this
work (Jayaratne et al., 2018). Both effects can be reduced significantly by drying the air.
Therefore, in future work, we intend to investigate using a sensor augmentation with a
simple low-cost heated air inlet to compensate (Fig. 6b).
Besides humidity effects, we expect age related deviations in long-term use due to the
design of the sensor (e.g. due to residual dust in the sensor and aging of the fan). We are
currently investigating this in a long-term real-world deployment. Finally, with non-expert
users, an often overlooked source of error is the human operator (Budde et al., 2017b).
(a) (b)
Fig. 6. The SDS011 is very susceptible to humidity. In fog, e.g., readings overestimate the actual particle
concentration drastically (a). In future work, we intend to investigate drying the air using a low-cost do-it-yourself
(DIY) sensor augmentation with a heated air inlet (b) to compensate this effect
1.1 Polydisperse Particulates
We compared several SDS011 sensors with reference measurements in a lab setting at
the World Calibration Center for Aerosol Physics (WCCAP) at the Leibnitz Institute for
Tropospheric Research (TROPOS) that is operated in cooperation with the German Federal
Environmental Agency (UBA) and the World Meteorological Organization (WMO). We
exposed a set of 17 SDS011 sensors to different aerosols (ammonium sulfate, ambient air,
soot and zero air) in an airtight aluminum chamber, depicted in Figure 3.
Fig. 3. SDS011 sensors were placed inside an aluminum measurement chamber (left), into which a varying
concentration of polydisperse particles was injected. An SMPS and a TSI APS (right) were used as reference.
The air outlet was connected to two reference devices: A TSI APS and an SMPS custom
made at the WCCAP (Wiedensohler et al., 2012). This combination enabled the
measurement of 92 aerodynamic size channels between 10nm and 20µm. Time resolution
was one reading every 4.5 minutes. The SMPS samples the different channels in a time
multiplex fashion, i.e. one after another over of 4.5 minutes. From this, we calculated the
three size classes PM10, PM2.5 and PM1.
Fig. 4. Time series of experiments with different polydisperse aerosols (left) and zoomed in for PM2.5 mass for
the first session of ambient air measurements (right). The reference is the solid black line.
Figure 4 shows the time series of our experiments. Three individual sensors were
disregarded, as was the complete ambient air 2 session, due to technical errors,. We can see
that the SDS011 generally captures dynamics well, but also that the readings have an offset
Ambient
Air 1
Ambient
Air 2
Zero
Air
Ammonium
Ammonium
Sulfate 2
Soot
10 ProScience 5 (2018) 6-12
Figure 4 shows the time series of our experiments. Three individual sensors were
disregarded, as was the complete ambient air 2 session, due to technical errors. We can see
that the SDS011 generally captures dynamics well, but also that the readings have an offset
to the reference and that this offset differs for the individual sensors. This variance can also
be seen in the scatterplot in Figure 5.
Fig. 4. Time series of experiments with different polydisperse aerosols (left) and zoomed in for PM2.5 mass for the
first session of ambient air measurements (right). The reference is the solid black line
Fig. 5. Scatterplots: mean PM2.5 mass against the reference (left) and for all individual 14 valid sensors for the test
aerosol ambient air (right).
We see that the sensor exhibits overall good linearity, but also systematic misestimation
of the true concentration. On average, the SDS011 showed only 66 % of the PM2.5 in
ambient air. Individually, the readings ranged from 45 % to 85 %. For ammonium sulfate,
the SDS011s on average show only 48 % of the reference’s PM2.5. For pure soot, no usable
readings could be obtained, as was expected due strong absorption and low light scattering.
This is a general issue of optical measurements and not specific to the SDS011. It should be
mentioned that soot in ambient air is coated with other materials, e.g. organics and sulfates,
and scatters light. A general conversion factor how much soot is seen by the SDS011
cannot be given. For zero air, the sensors showed minimum PM2.5 values of ~0.3 µg/m³,
which seems acceptable for ambient air measurements. All results are presented in Table 1.
2.3 Other Factors
The other most noteworthy aspect that affects the SDS011 sensor is humidity.
Generally, the sensors are only specified to operate up to a relative humidity of 70% (Nova
Fitness, 2015). Above these values, large deviations must be expected. Figure 6a shows one
Table 1. Measurement results for the individual sensors (PM2.5 / PM10).
Sensor
Ambient air
Ammonium sulfate
Soot
Zero air
slope
intercept
slope
intercept
slope
intercept
baseline
stdev
1188476 0.85 / 0.85 -0.88 / -0.74 0.63 / 0.67 0.01 / 0.26 0 / 0 1.55 / 1.84 0.247 / 0.454 0.0020 / 0.0018
1190049
0.67 / 0.66
-0.66 / -0.50
0.50 / 0.53
0.19 / 0.40
0 / 0
1.21 / 1.39
0.303 / 0.400
0.0009 / 0.0002
1197947
0.61 / 0.60
-0.68 / -0.59
0.46 / 0.49
0.23 / 0.63
0 / 0
1.09 / 1.37
0.200 / 0.300
0.0007 / 0.0004
1200580
0.84 / 0.84
-0.60 / -0.59
0.62 / 0.66
0.24 / 0.41
-0.01 / -0.01
1.95 / 2.09
0.415 / 0.413
0.0017 / 0.0024
132097
0.65 / 0.64
-0.79 / -0.57
0.47 / 0.50
0.17 / 0.72
0 / 0
1.11 / 1.38
0.200 / 0.402
0.0003 / 0.0024
1547350
0.60 / 0.59
-0.80 / -0.76
0.47 / 0.51
-0.06 / 0.55
0 / 0
1.09 / 1.24
0.301 / 0.400
0.0004 / 0.0002
1547430
0.69 / 0.65
-0.98 / -0.67
0.48 / 0.53
0.03 / 0.94
0 / 0
1.26 / 1.68
0.300 / 0.800
0.0000 / 0.0003
1547436
0.64 / 0.60
-0.80 / -0.44
0.51 / 0.56
-0.12 / 0.34
0 / 0
1.27 / 1.70
0.393 / 0.700
0.0014 / 0.0000
1548172
0.48 / 0.45
-0.53 / 0.12
0.42 / 0.45
0.04 / 1.44
0 / 0
0.97 / 1.82
0.300 / 0.600
0.0002 / 0.0005
1548187
0.65 / 0.71
-0.52 / -0.48
0.47 / 0.53
0.36 / 1.08
0 / 0
1.28 / 1.84
0.407 / 0.601
0.0010 / 0.0008
1548303
0.66 / 0.69
-0.59 / -0.69
0.50 / 0.55
0.11 / 0.55
0 / 0
1.26 / 1.49
0.391 / 0.500
0.0012 / 0.0002
1581752
0.64 / 0.62
-0.62 / -0.41
0.396 / 0.699
0.0017 / 0.0013
158693
0.63 / 0.62
-0.76 / -0.62
0.46 / 0.49
0.21 / 0.53
0 / 0
1.08 / 1.37
0.200 / 0.401
0.0002 / 0.0004
1590613
0.52 / 0.52
-0.73 / -0.52
0.41 / 0.45
-0.13 / 0.65
0 / 0
0.94 / 1.15
0.300 / 0.400
0.0000 / 0.0004
Average
0.65 / 0.65
-0.71 / -0.53
0.49 / 0.53
0.10 / 0.65
day of data collected in-the-wild on a humid day. The graph shows that the sensor
drastically overestimates the particle concentration in the morning and evening hours when
humidity is high. There are two sources of error caused by humidity: Hygroscopic growth
of saline particles and condensed fog droplets of similar size as PM-relevant particles. We
observed that the overestimation is much stronger in fog, and less extreme in otherwise
high humidity or rain. These findings are in line with results published in parallel with this
work (Jayaratne et al., 2018). Both effects can be reduced significantly by drying the air.
Therefore, in future work, we intend to investigate using a sensor augmentation with a
simple low-cost heated air inlet to compensate (Fig. 6b).
Besides humidity effects, we expect age related deviations in long-term use due to the
design of the sensor (e.g. due to residual dust in the sensor and aging of the fan). We are
currently investigating this in a long-term real-world deployment. Finally, with non-expert
users, an often overlooked source of error is the human operator (Budde et al., 2017b).
(a) (b)
Fig. 6. The SDS011 is very susceptible to humidity. In fog, e.g., readings overestimate the actual particle
concentration drastically (a). In future work, we intend to investigate drying the air using a low-cost do-it-yourself
(DIY) sensor augmentation with a heated air inlet (b) to compensate this effect
1.1 Polydisperse Particulates
We compared several SDS011 sensors with reference measurements in a lab setting at
the World Calibration Center for Aerosol Physics (WCCAP) at the Leibnitz Institute for
Tropospheric Research (TROPOS) that is operated in cooperation with the German Federal
Environmental Agency (UBA) and the World Meteorological Organization (WMO). We
exposed a set of 17 SDS011 sensors to different aerosols (ammonium sulfate, ambient air,
soot and zero air) in an airtight aluminum chamber, depicted in Figure 3.
Fig. 3. SDS011 sensors were placed inside an aluminum measurement chamber (left), into which a varying
concentration of polydisperse particles was injected. An SMPS and a TSI APS (right) were used as reference.
The air outlet was connected to two reference devices: A TSI APS and an SMPS custom
made at the WCCAP (Wiedensohler et al., 2012). This combination enabled the
measurement of 92 aerodynamic size channels between 10nm and 20µm. Time resolution
was one reading every 4.5 minutes. The SMPS samples the different channels in a time
multiplex fashion, i.e. one after another over of 4.5 minutes. From this, we calculated the
three size classes PM10, PM2.5 and PM1.
Fig. 4. Time series of experiments with different polydisperse aerosols (left) and zoomed in for PM2.5 mass for
the first session of ambient air measurements (right). The reference is the solid black line.
Figure 4 shows the time series of our experiments. Three individual sensors were
disregarded, as was the complete ambient air 2 session, due to technical errors,. We can see
that the SDS011 generally captures dynamics well, but also that the readings have an offset
Ambient
Air 1
Ambient
Air 2
Zero
Air
Ammonium
Ammonium
Sulfate 2
Soot
ProScience 5 (2018) 6-12 11
3. Conclusions
In summary, our experiments show that:
The SDS011 can capture dynamics of fine dust with high temporal resolution.
There is a notable variance between individual sensors.
PM2.5 readings seem promising, esp. for background PM sensing.
PM10 estimates may be wrong, esp. if distribution shifts towards larger particles.
Humidity dependence is a problem, esp. in fog.
Without further measures, the sensor is only appropriate for qualitative not
quantitative data.
The SDS011 is less sensitive to ambient soot.
That being said, the sensor shows remarkable stability for its cost and has the potential
to enable novel applications. In future work, we will explore humidity compensation
methods, e.g. sensor augmentation, sensor data fusion, and networked sensing scenarios of
low-cost sensors combined with big data analytics approaches (Budde et al. 2017). We will
also report on the long-term stability of the sensors in a real-world deployment.
4. Acknowledgements
Partial funding: BMBF project "Software Campus" (01IS12051) and BMVI project "SmartAQnet" (19F2003B).
References
Budde M., El Masri R., Riedel T., Beigl M. (2013). Enabling Low-Cost Particulate Matter Measurement for
Participatory Sensing Scenarios, 12th Int. Conference on Mobile and Ubiquitous Multimedia (MUM 2013).
Budde M., Köpke M., Beigl M. (2015). Robust, In-situ Data Reconstruction from Poisson Noise for Low-cost,
Mobile, Non-Expert Environmental Sensing. International Symposium on Wearable Computing (ISWC’15).
Budde M., Riedel T., Beigl M., Schäfer K., Emeis S., Cyrys J., Schnelle-Kreis J., Philipp A., Ziegler V., Grimm
H., Gratza T. (2017). SmartAQnet: Remote and In-Situ Sensing of Urban Air Quality, Proc. SPIE 10424,
Remote Sensing of Clouds and the Atmosphere XXII, 104240C.
Budde M., Schankin A., Hoffmann J., Danz M., Riedel T., Beigl M. (2017b). Participatory Sensing or
Participatory Nonsense? – Mitigating the Effect of Human Error on Data Quality in Citizen Science.
Interactive, Mobile, Wearable and Ubiquitous Technology (IMWUT) 1(3).
Holstius D.M., Pillarisetti A., Smith K.R., Seto E. (2014). Field calibrations of a low-cost aerosol sensor at a
regulatory monitoring site in California. Atmospheric Measurement Techniques 7.
Jayaratne R., Liu X., Thai P., Dunbabin M., Morawska L. (2018). The influence of humidity on the performance
of a low-cost air particle mass sensor and the effect of atmospheric fog. Atmospheric Measurement Tech. 11.
LUBW (2017). Messungen mit dem Feinstaubsensor SDS011 - Ein Vergleich mit einem eignungsgeprüften
Feinstaubanalysator. https://www4.lubw.baden-wuerttemberg.de/servlet/is/268831/ (accessed 01/24/2018).
Nova Fitness Co., Ltd. (2015). SDS011 sensor. V1.3, https://nettigo.pl/attachments/398 (accessed 01/24/2018).
Schwarz A.D., Meyer J., Dittler A. (2018). Opportunities for Low-Cost Particulate Matter Sensors in Filter
Emission Measurements. Chemical Engineering & Technology 41(9).
Snyder E.G., Watkins T.H., Solomon P.A., Thoma E.D., Williams R.W., Hagler G.S.W., Shelow D., Hindin D.A.,
Kilaru V.J., Preuss P.W. (2013). The changing paradigm of air pollution monitoring. Environmental Science
& Technology 47(20).
Wiedensohler A., Birmili W., Nowak A., Sonntag A., Weinhold K., Merkel M., Wehner B., Tuch T., Pfeifer S.,
Fiebig M., Fjäraa A.M., Asmi E., Sellegri K., Depuy R., Venzac H., Villani P., Laj P., Aalto P., Ogren J.A.,
Swietlicki E., Williams P., Roldin P., Quincey P., Hüglin C., Fierz-Schmidhauser R., Gysel M., Weingartner
E., Riccobono F., Santos S., Grüning C., Faloon K., Beddows D., Harrison R., Monahan C., Jennings S.G.,
O’Dowd C.D., Marinoni A., Horn H.-G., Keck L., Jiang J., Scheckman J., McMurry P.H., Deng Z., Zhao C.S.,
Moerman M., Henzing B., de Leeuw G., Löschau G., Bastian S. (2012). Mobility particle size spectrometers:
Harmonization of technical standards and data structure to facilitate high quality long-term observations of
atmospheric particle number size distributions, “Atmospheric Measurement Techniques” 5(3).
12 ProScience 5 (2018) 6-12
