Effect of an Aerosol Dryer on Ambient PM Measurements with SDS011 Low Cost Sensors during a Two-year Period in Duisburg, Germany

The performance of low-cost particulate matter (PM) sensors (NovaFitness SDS011) in connection with a self-developed aerosol dryer has been investigated in ambient air measurements over a two-year period by comparing the reported mass concentration of particulate matter (PM x ) with the results of a PM 10 reference filter sampler and two Tapered Element Oscillating Microbalances (TEOM), one for PM 10 and one for PM 2.5 . Special emphasis was put on the effect of relative air humidity on sensor readings. In total, four sensors were used per year in two pairs. For one pair, the aerosol was dried with a newly developed low-cost aerosol dryer, whereas the other pair measured the untreated aerosol. The results show that the 24 h average concentration reported by the sensor could be by a factor of up to 38 too high compared to the gravimetric measurement, if the aerosol was not dried, whereas the mismatch with a maximum factor of 5.4 was significantly lower when using the dryer. For the PM 10 concentration correction factors were determined from the ratios of the measured 24 h values of the sensors and the gravimetric reference. Corrected 24 h mean PM 10 concentrations with dryer agreed mostly within a factor of 2 with the gravimetric reference data from the filter sampler, whereas results from measurements without dryer agreed only within a factor of 5. The results further show that the sensors underestimate the mass concentrations in case of low relative humidity ( < 45%). Similar results are seen for the PM 2.5 data as well. Therefore, the use of a constant correction factor was checked. It is shown that an average correction factor of around 2.5 for the PM 10 and around 2.6 for the PM 2.5 readings could be a reasonable approach for all SDS011 sensors equipped with the newly developed aerosol dryer.


INTRODUCTION
Air pollution is known to cause adverse effects on public health.In 2018, the Lancet commission concluded that air pollution causes 6.5 million premature deaths worldwide every year (Landrigan et al., 2018) and ranked it as fifth leading cause of death (Forouzanfar et al., 2015).According to Lelieveld et al. (2020), even 8.8 million people died prematurely in 2015 due to increased air pollution, corresponding to a reduced life expectancy of 2.9 years.The effect of air pollution on life expectancy was thus higher than the effect of smoking (2.2 years).The authors further explain that 5.5 million premature deaths would be avoidable because a large fraction of the worldwide air pollution stems from burning of fossil fuels.Particulate matter (PM) is one of the major air pollutants and correlations between increased PM concentrations and adverse health effects have been reported in a multitude of epidemiological and toxicological studies (Dockery et al., 1993;Dockery, 2009;Pope III and Dockery, 2006;Rückerl et al., 2011;Lelieveld et al., 2015).Consequently, many countries and regions have issued limit values for PM2.5 and PM10, i.e., for mass concentrations of particles with 50% aerodynamic cut off diameters of 2.5 µm and 10 µm, respectively.For example, in the European Union (EU) as per Council Directive 1999/30/EC, the 24 h limit value for PM10 is 50 µg m -3 which must not be exceeded on more than 35 days per calendar year.The annual PM10 limit value is 40 µg m -3 (EU, 1999).For PM2.5, only an annual limit value of 25 µg m -3 exists in the EU and it only needs to be controlled in the urban background, but not at traffic or industry sites (EU, 2008).While in the EU the main focus is on PM10, other countries like the USA, China or India put more emphasis on PM2.5 and have corresponding limit values in place which in some cases are stricter than the ones in the EU.In 2021, the World Health Organization (WHO), whose guidance values for PM concentrations have always been well below the European limit values, further reduced their suggested PM10 limits to 45 µg m -3 for 24 h means and 15 µg m -3 annual means.The corresponding values for PM2.5 are 15 µg m -3 and 5 µg m -3 , respectively.In the wake of the new WHO recommendations, the EU has announced that it will reduce its limits as well.
The reference method to determine PM concentrations is described e.g., in EN 12341:2014(EN, 2014) and comprises a size-selective inlet, usually an impactor, that removes particles which are too large to comply with the PM2.5 and PM10 convention, respectively.All remaining particles are collected on a sample filter at a constant air flow rate.The mass gain of the filter is determined by pre-and post-weighing of the filter and the mean mass concentration calculated by dividing the difference of the two weighings by the volume of air sampled.The filter can further be used, e.g., for chemical analyses of the collected dust.While the method is very simple and robust, it only provides limited temporal information (24 h averages) on the particle concentration.Alternative measurement methods (also termed automatic methods), which have proven their equivalence to the reference method, include the Tapered Element Oscillating Microbalance (TEOM, Thermo Fisher (Patashnick and Rupprecht, 1991)), Beta Attenuation Monitors (various manufacturers, (Macias and Husar, 1976)) and optical aerosol spectrometers (various manufacturers (Spielvogel and Weiss, 2013)), which all provide much higher time resolution of at least a few minutes.Thus, they allow for the determination of diurnal variations of local PM concentrations.Many atmospheric measurement stations for regulatory purposes are equipped with a combination of filter and automatic methods to obtain both high temporal resolution as well as information on the chemical particle composition by extracting and/or digesting chunks of the filter samples, followed by wetchemical analyses, e.g., ion chromatography for anions, inductively coupled plasma optical emission spectrometry (ICP-OES) or inductively coupled plasma mass spectrometry (ICP-MS) for cations or photometry for ammonia.Such stations are usually part of measurement networks that contain measurement locations of different characteristics, e.g., regional background, rural background, urban background and hot spots such as traffic sites (Lenschow et al., 2001).While a suitable choice of measurement locations according to these classifications allows for at least a rough differentiation of possible particle sources, the high costs of the measurement equipment prevent the establishment of dense measurement networks.As an alternative, a variety of low-cost PM sensors has entered the market and become very popular over the recent years.Their prices are mostly well below 100 €.These sensors detect the light scattered by particles, drawn through an optical measurement volume and determine the particle mass concentration based on a manufacturer calibration.Most sensors deliver the mass concentration of at least two size fractions, i.e., PM2.5 and PM10.Consequently, besides the measurement of the total light scattered by a cloud of particles, they must also obtain some information on particle sizes, e.g., by a combination of a photometric measurement with single particle pulse height analysis (Wang et al., 2009).However, most sensor manufacturers neither disclose the exact measurement principle nor their calibration procedure.
Due to their low price, the use of these sensors for establishing dense measurement networks has been proposed (Clements et al., 2017;Sousan et al., 2018;Gao et al., 2015;Giordano et al., 2021).In 2015, a large citizen science project was initiated in Stuttgart (Germany), a city known for its increased air pollution.The organizers provide instructions for building a PM measurement device based on the low cost PM sensor module SDS011 (NovaFitness), which is available for around 30 €.It reports PM2.5 and PM10 concentrations with up to 1 s time resolution.The final device includes a microcontroller (NodeMCU) to retrieve the PM data and a WiFi module to upload data to a cloud when connected to a local network.The organizers further suggest the use of a low-cost temperature and humidity sensor (DHT22, Aosong Electronics) and two connected waste water pipe elbows as a weatherproof housing.The total price of the final device is around 50 €.The measured PM2.5 and PM10 data from all participating sensors are displayed in a color-code on a map on the initiative's website (https://sensor.community/en/).In the meantime, the sensor network has expanded over large parts of Europe and contains several thousand individual sensors.Whereas the sheer number of sensors in the network provides a high spatiotemporal resolution for the PM concentrations, the accuracy of the data has been questioned, particularly due to the lack of quality assurance and because of known air humidity effects on optical aerosol measurements.
It is well known that hygroscopic particles take up water when exposed to high relative humidity and thus grow (Köhler, 1936).Consequently, hygroscopic particles scatter more light with increasing humidity (MacKinnon, 1969), which is misinterpreted by the optical aerosol measurement devices as a higher PM concentration.Optical aerosol spectrometers, which are approved for regulatory control of atmospheric PM concentrations, therefore use aerosol dryers to assure that the particles do not contain water.These dryers are, however, not useable with low-cost sensors, because they introduce an additional pressure drop, which cannot be overcome by the built-in fans of the sensors.Furthermore, the available aerosol dryers cost a multiple of the low-cost sensors and are therefore not economically viable for this application.A retrospective data correction of photometric measurements has been proposed (Hua et al., 2021;Molnár et al., 2020;Soneja et al., 2014;Hofman et al., 2022), but such a correction suffers from the lack of an unambiguous relation of the size of hygroscopic particles and relative humidity.Instead, the hygroscopic growth shows a hysteresis, i.e., the humidity level at which particles start to grow during increasing humidity (deliquescence) is higher than the humidity level at which they have completely released their water content during decreasing humidity (efflorescence).Furthermore, the hygroscopic growth depends on the particle material and thus on the chemical composition of the PM (Hansson et al., 1998;McInnes et al., 1998).Typical hygroscopic particle materials in the atmosphere include e.g., ammonium nitrate (NH4NO3), ammonium sulfate ((NH4)2SO4) or sodium chloride (NaCl).
Several researchers have reported adverse effects of relative humidity on measurements with low-cost PM sensors (Wang et al., 2021;Zou et al., 2021;Bulot et al., 2020;Stavroulas et al., 2020;Tryner et al., 2020;Brattich et al., 2020).Jayarathne et al. (2018) used a Plantower PMS1003 low-cost sensor at a measurement site in Brisbane, Australia for 24 days and found the sensor to report significantly increased concentrations at humidity levels above 75%.Crilley et al. (2018) compared the performance of 14 Alphasense OPC-N2 low cost aerosol spectrometers against reference instruments at a measurement site in Birmingham, UK.They found that particles significantly grew and that the PM concentrations reported by the OPC-N2 instruments were hence over-estimated at relative humidity levels above 65%.Tagle et al. (2020) used NovaFitness SDS011 sensors to monitor the PM2.5 and PM10 concentrations at three different sites in Santiago, Chile and compared the data to reference data from a beta attenuation monitor (PM2.5) and TEOM (PM10).They found that the 24 h average PM2.5 data from the sensors were in reasonable agreement with the reference data, but also observed considerable bias of the 1 h average concentrations.In particular, the authors report about significant overestimations of the PM concentrations at relative humidity levels above 75% and underestimations at humidity levels below 50%.Liu et al. (2019) investigated the performance of three NovaFitness SDS011 sensors over a four-month period at a measurement site in Oslo, Norway.They only evaluated the PM2.5 concentrations reported by the sensor and compared it with a TEOM with filter dynamics measurement system (FDMS (Grover et al., 2005)).They found that the sensor data showed reasonably good linearity against the TEOM data but also report a high overestimation of the PM2.5 concentrations at high relative humidity (> 80%).Božilov et al. (2022) found, that high relative humidity (> 70%) especially affected the PM10 measurement of the SDS011 sensors negatively.Several publications (Asbach et al., 2018;Masic et al., 2020;Budde et al., 2018) therefore recommend the development and use of a simple, low-cost dryer for improving data quality of low-cost PM sensors.While such a dryer has to efficiently remove humidity from the air, it must not alter the particle size, shape or concentration.Masic et al. (2020) compared a high-end spectrometer 11-D and SMPS from Grimm with two low-cost sensors (Alphasense OPC-N2 and MAQS, equipped with a Plantower PMS5003 (Jiang et al., 2011)) and used a diffusion dryer to dry the aerosol upstream of the 11-D spectrometer and the SMPS, but not the low-cost sensors.They suggested to further investigate the effect of micro-dryers specifically designed for low-cost sensors.While a silica gel based diffusion dryer dries the aerosol effectively during a short term measurement, its efficacy is imparted over time due to the water loading of the silica gel, requiring it to be frequently exchanged and regenerated.Chacón-Mateos et al. (2022) developed a thermal low-cost dryer and tested it in laboratory experiments with an Alphasense OPC-R1 low cost spectrometer.They found that the dryer can effectively reduce the influence of high relative humidity (70-90%) on the PM2.5 and PM10 measurement, but struggled with the discontinuity of the drying process of the developed dryer.They also highlighted the need for a sensor calibration in addition to the use of the dryer.Samad et al. (2021) tested a self-developed thermal dryer, that operates at low voltage, with an Alphasense OPC-N3 sensor under laboratory conditions in order to find the optimal settings of the heating unit.They concluded that the aerosol dryer is able to eliminate the negative effects of relative humidity on the measured values of the OPC-N3 sensor.However, a low cost aerosol dryer with a self-regulating heating wire which can be operated at 230 V seems to be more user friendly for private use.
The aim of this study was to investigate the performance of NovaFitness SDS011 sensors in atmospheric measurements over an extended period of time and to investigate the effect of relative humidity on the results in comparison to reference data.A simple, low cost aerosol dryer was developed and evaluated in this study in order to evaluate the potential of aerosol drying to improve the data quality of low-cost sensors (Nothhelfer, 2020).

MEASUREMENT LOCATION AND EQUIPMENT
Measurements have been conducted on the parking lot of the Institute of Energy and Environmental Technology (IUTA) in Duisburg, Germany (GPS coordinates 51°23′21′′N, 6°43′33′′E) between March 1 st , 2020 and March 1 st , 2022.The institute is located in the midst of a logistics center of Europe's largest inland port.Local aerosol sources during weekdays are therefore dominated by traffic, especially of heavy duty vehicles, whereas the local traffic is very low at night and during weekends.Vehicle movements on the institute's parking lot mainly occur in the mornings and afternoons on weekdays, but are rather low otherwise.The measurement location is at the very western end of the Ruhr area.During the most typical north-westerly winds, the air masses come from mainly rural areas, whereas during the less frequent easterly winds, the air masses have passed the Ruhr area with more than 5 million inhabitants.The measurement location is not directly surrounded by larger buildings and is thus exposed to free wind flows.Accumulation of dust, like often observed in street canyons, can therefore be excluded.The location can thus not be clearly categorized according to Lenschow et al. (2001), but would fall in-between an urban background and a traffic site.In order to keep track of meteorological conditions, the measurement station was equipped with a weather station (Vaisala WXT 510) that measures temperature, relative humidity, wind direction and wind speed.Additionally, the measurement station was equipped with a heating and cooling unit to constantly hold an inside temperature of around 20°C.
The measurement station contained two TEOMs (Thermo Fisher, model 1400 AB (Patashnick and Rupprecht, 1991)).One device was equipped with a Rupprecht & Patashnick standard PM10 impactor (model number 57-000596) to measure the PM10 concentration.The second device was additionally equipped with a PM2.5 sharp-cut cyclone (Rupprecht & Patashnick model number 57-005896) downstream of the PM10 impactor to measure the PM2.5 concentration.The inlets of both TEOMs are operated at a flow rate of 1 m 3 h -1 (16.67 L min -1 ).The TEOMs were set to a temperature of 40°C to lower the relative humidity in the measurement compartment.40°C was chosen instead of the usual temperature of 50°C to minimize losses of semivolatile particles (Allen et al., 1997;Meyer et al., 2000).FDMS systems were not available and thus losses of semi-volatile particles, but especially insufficient drying of the particles, still have affected the TEOM results (Li et al., 2012).While an optical aerosol instrument measures the light scattered by particles passing through the internal measurement volume in real time, the TEOM collects the particles on a filter and continuously measures the rate at which the mass of the filter changes.During "normal" operation, the filter mass constantly increases, due to the collection of particles.However, this mass gain is superimposed by a mass loss, if the sampled aerosol contains (semi-) volatile particles and water content.Therefore, from time to time and depending on the temporal evolution of the chemical composition of the particles, it sometimes happens that the mass loss of the filter is higher than the mass gain, which is (mis-)interpreted by the TEOM as a negative concentration.Due to the reduced heating temperature upstream of the TEOM the largest influence of the result is most likely non-evaporated water content.PM10 data from the TEOM was recorded with 1 min time resolution and later averaged to obtain 1 h and 24 h mean values.Negative 1 h and 24 h averages, which may result from the loss of semivolatile particles or by a humidity effect on the measurement, were excluded from further analysis.
The measurement station further contained a low-volume sampler (Derenda LVS), equipped with a PM10 impactor and an automatic filter changer.The LVS sampled at a flow rate of 2.3 m 3 h -1 on 47 mm quartz fiber filters.The filter changer was programmed to automatically change filters at midnight (central European winter time) in order to obtain 24 h daily averages of the PM10 concentration.PM10 was preferred over PM2.5 as it is the quantity that has to be assessed in European air quality monitoring.To also derive a calibration factor for PM2.5, a filter sampler with an upstream PM2.5 impactor was operated in parallel to the PM2.5 TEOM measurement for 10 days.As no second LVS system was available for the two-year period, the focus of this paper lies on the PM10 measurement.
The filters were handled according to EN 12341:2014, i.e., they were equilibrated in a weighing room to a temperature of (20 ± 1)°C and a relative humidity of (47.5 ± 2.5)% (1 h averages) for at least 48 h prior to weighing before and after sampling.Each filter was weighed twice with at least 24 h in-between for each pre-and post-weighing procedure.The determined filter masses were only deemed valid if the discrepancy of both weighings was less than 40 µg (before) and 60 µg (after sampling), respectively.If the difference was larger, the filter was weighed for a third time after another 24 h and the data only used if the discrepancy was now within the aforementioned range.Otherwise, the data for this filter were omitted.This procedure also guarantees a minimized losses of semi-volatile particles.The masses of lab and field blank filters were recorded as a quality control measure and were all within the limits set in the standard.The determined PM10 daily mean values were then used to check and, if necessary, to correct the calibration of the TEOM and the SDS sensors with upstream aerosol drying.
A total of four SDS011 PM sensors were installed at the measurement station.A pair of two sensors were each equipped with a homemade low-cost heating unit to dry the aerosol before entering the SDS011 sensors.Although the heating unit does not dry the air due to a constant total humidity, it still dries the airborne particles by reducing the relative humidity and is therefore called dryer in the following.As shown in Fig. 1, the aerosol dryer consists of a copper tube wrapped in a self-regulating heating wire.The self-regulating heating wire has a theoretical maximum temperature of 65°C (Raychem, 2023).With increasing temperature of the self-regulating heating wire, its resistance increases and thus the heating power decreases.Consequently, no further electrical control for the heating unit is necessary.In addition, the aerosol dryer was thermally insulated and the complete measuring unit placed inside a waterproof housing.In order to monitor the functionality of the heating unit, an additional temperature sensor was placed between heater and insulation.
Prior laboratory tests with outdoor temperatures of 10°C to 40°C have shown, that this heating unit dries the aerosol efficiently to less than 45% relative humidity up to an outdoor temperature of 32°C and a relative humidity of 100%.At higher outdoor temperatures the heating unit is not able to reduce the relative humidity from 100% to under 50% but might still be sufficient at lower relative humidity levels.Additionally, a possible effect of the increased aerosol temperature on the sensor performance has been investigated in prior laboratory experiments, but was found to be negligible (Nothhelfer, 2020).All prior laboratory tests can be seen more detailed in (Nothhelfer, 2020).The temperature in Germany rarely exceeds this range at such high humidity levels.Consequently, the heating unit was considered sufficient for this measurement location.The two SDS011 PM sensors equipped with heating unit are denominated as "SDS-heated-1" and "SDSheated-2" in the following.The lowest temperature, at which the heating unit can be operated, has not yet been determined.Especially temperatures below 0°C could lead to an insufficient heating state, due to frozen water content of the particles.
The other two sensors were set up in exactly the same fashion as proposed by the citizen science initiative from Stuttgart (see https://sensor.community/en/sensors/airrohr/ for more  (Nothhelfer, 2020).information), with the exception that the data was not automatically uploaded to their website, but stored on a local PC for subsequent processing.This sensor setup also included the temperature and humidity sensors (DHT22, Aosong Electronics).Both sensor setups were installed on the roof of the measurement container in weatherproof housings and approximately 50 cm apart from each other.They are denominated "SDS-3" and "SDS-4" in the following.PM2.5, PM10, temperature and humidity were recorded with 1 s time resolution and later averaged over 1 h and 24 h, respectively, to obtain the same time resolution as the TEOM and LVS data, respectively.
After approximately one year of measurement, the installed SDS011 sensors were replaced by new SDS011 sensors, since the manufacturer specifies a maximum runtime of 8000 hours for continuous operation of the sensors (Nova Fitness Co., Ltd., 2015).

Overview of the Time Series and Data Validation
The time series of the measured PM2.5 and PM10 concentrations as well as temperature and relative humidity are shown in Fig. 2 as 24 h averages for the whole measurement period of two years.The time of the sensor exchange is marked with a red line in the graph.Average values were only deemed valid if at least two thirds of the 1-minute data points from the averaging period were available.In case of the TEOM, invalid data mostly stemmed from either negative concentration values due to evaporation of semivolatile particles or water from the sample filter, or from too high noise levels, e.g., after filter changes (approximately once per month).Invalid SDS011 sensor data mostly arose from a loss of the connection between the sensors and the PC due to loose USB-connections.The DHT22 sensors and temperature sensors at the heating unit failed after only a few weeks, and are therefore not used in the further evaluation.Temperature and humidity data from the weather station are used instead, where applicable.Occasional short outages of the SDS011 sensors and the weather station were additionally caused by PC crashes.
The overall data coverage of the individual devices is listed in Table 1.For the hourly data, the data coverage of the SDS011 sensors was around 85% and thus only slightly below the data coverage of the two TEOMs.Since the loss of connections and negative concentration readings of the TEOMs typically did not last very long, the coverage of the 24 h data was much better, i.e., of around 95.5% in case of both TEOMs.The uncorrected daily mean values of the SDS011 sensors are shown in Fig. 2. Since the gravimetric measurement was only available for the PM10 concentration, it is used as reference in (b), whereas corrected TEOM PM2.5 values according to Eq. ( 1) are used as reference for the PM2.5 concentration (c).Looking at the time series of the pair of SDS sensors with heating unit (insert at (b)), it is noticeable that after about two months of measurement, the SDS-heated-2 sensor (orange) suddenly showed permanently lower values than the SDS-heated-1 sensor (red), especially for PM10.After the sensors were exchanged on February 26, 2021, this behavior can no longer be seen.Although the mismatch between the results of the two sensors had been noticed after two months, it had been decided to continue the measurements with the same sensors in order to obtain long term data with identical sensors.

Derivation of Correction Factors
The concentrations measured by the two sensors without heating unit (blue colors) show high peaks ostensibly random during some periods and an underestimation during others (see Fig. 2).On the other hand, the graph also shows that the sensors with dryer underestimated the reference concentrations almost the entire time.Apparently, the SDS011 sensors tend to underestimate PM concentrations at low relative humidity levels (< 45%).Chacón-Mateos et al. (2022) found similar results for a low-cost PM-sensor (Alphasense OPC-R1) when drying the aerosol upstream in an ambient measurement campaign in Stuttgart.Although the sensor manufacturers do not disclose their calibration protocols, it is likely that they calibrate the sensors to provide most accurate results for average relative humidity levels around 50%.It should be noted that also the TEOM tends to underestimate mass concentrations due to the heating of the sample filter.Ayers et al. (1999) suggested the loss of semi-volatile particles from the heated sample filter as the root-cause for this behavior.To correct for this, a correction factor according to Eq. ( 1) was calculated following the common practice for TEOM operation at official measurement stations.
The correction factor Fs is formed from the average ratio of the daily mean PMx concentration of the reference measurement (e.g., filter measurement for PM10) to the PMx concentration of the sensor to be corrected.Here subscript r stands for reference and s for sensor.To obtain the corrected results, the measured values of the sensors are multiplied with the correction factor Fs.This correction is done for the TEOM and the SDS011 sensors with heating unit.For the comparison of the PM2.5-and hourly mean values of the PM10 concentrations, the corrected TEOM values are used as a reference.Although the low-cost heating unit, developed here, also heats the aerosol, substantial evaporation of semivolatile particles can be excluded due to the short residence time of the aerosol in the heated atmosphere.
For widespread application of low-cost PM sensors with upstream aerosol drying, it is necessary to keep the entire measuring system as low-maintenance as possible.Consequently, only a constant correction factor is feasible, since frequent recalibrations will-if at all possible-likely not be accomplished by private users.For the SDS sensors, the question is whether a constant correction factor could be applied for the long-term use of every SDS011 sensor.To check whether a constant correction factor is suitable for different sensors, we looked at the variation of the correction factor with time.Fig. 3 shows the correction factors, calculated according to Eq. ( 1) for the individual months of the measurement period.Comparing the curves of the different measuring instruments, the correction factor of the TEOM, shown in black, shows the lowest variability.As the TEOM is one of the officially approved device for particulate matter monitoring, the standard deviation is only 9.9% for the first and 15.2% for the second year (see Table 2).Such a reference method must of course demonstrate high quality criteria, which may lead to low scattering of the measurement results.
As mentioned before, the offset among the first pair of SDS-sensors with dryer was increased between May 2020 and February 2021.This period is represented in Fig. 3 by the bold orange curve.It can be seen that the correction factor differs substantially during this time.After exchanging the SDS011 sensors at the end of February, the curves are again running parallel to each other.Therefore, this difference was obviously caused by a malfunction of the SDS-heated-2 sensor.Due to the malfunction of SDS-heated-2 during most of the first year, these data were disregarded for the determination of a constant correction factor and further analysis in general.
As expected, the correction factors are greatest for the sensors with heating.However, by looking at Table 2, the variability of the factors with and without heating are almost similar of around 30%, but it should be noted that the correction factors for the sensors with heating are constantly > 1, whereas those for the sensors without heating were sometimes also < 1.It appears that the SDS-sensors with and without heating may show a seasonal dependence.However, measurements over a longer period are needed for a robust statistical analysis for this observation.While the correction factor is highest during summer (June-August) the lowest factor in both years is around the turn of the year (December-January).Looking at the curve of the relative humidity in Fig. 3(a), it can be seen that the relative humidity is lower overall during summer than around the turn of the year which is contrary to the correction factors.Consequently, hygroscopic particle growth can be expected to be stronger during winter compared with summer time.Therefore, the SDS011 sensors without heating unit tend to overestimate (Fs < 1) the PM10 and PM2.5 concentrations in winter.Since this trend can also be seen with the SDS sensors with heating unit, it indicates that the drying efficiency may be insufficient during the winter months.The dependence of the daily average ratios of the corrected PM10 concentrations, measured with different sensors and TEOM, and the gravimetric reference concentration on the respective relative humidity range can be seen in Fig. 4. The data are corrected with the correction factors calculated according to Eq. ( 1) for the first 12 months and this correction factor is then used for all of the data.Since the data from SDS-heated-2 during the first year were omitted, the correction factor of SDS-heated-1, determined during the first year, was applied for the SDS-heated-2 sensor in the second year.Subsequently, the ratio of the corrected 24 h mean values of the sensors and the TEOM to the reference measurement was determined and plotted against the relative humidity.It should be noted that the ratios are plotted on logarithmic scales.
The whisker-boxes (upper and lower quartile) clearly show the strong humidity dependence of the ratio for the sensors installed without heating unit for humidity levels above approximately Fig. 4. Whisker-box-plot of the ratio of the 24 h average PM10 concentration, measured with SDS sensors and TEOM after applying the correction factor, and the gravimetric reference as a function of relative humidity.The correction factor of the first year was used for the second year as well; dashed reference lines at 0.5, 1, and 2. 70% and below approximately 40-50%, whereas the humidity dependence of the two sensors with aerosol dryer is much less pronounced.However, the constant drying of the particles by the heating unit prior to sensor entry generates an underestimation of the mass concentration.This reduces the dependence of the measurement results on the relative humidity, which means that the measured values of the SDS sensors with drying can be corrected more effectively by a simple correction factor than the SDS sensors without drying.
The whisker-boxes further show that the corrected data of the SDS011 sensors with dryer are more consistent with the filter measurement than the SDS sensors without heating unit.The ratios are almost all in a range from 0.5 to 2, i.e., the sensor data agree within a factor of 2 with the reference data.Also, the dependence of relative humidity almost vanished.

Determination of a Constant Correction Factor
As mentioned, for simplicity of handling for private users of the SDS011 sensors with dryer, it is desirable to be able to use a constant correction factor.For such a constant correction factor to be valid, the individual correction factors of the tested sensors should show a deviation as small as possible from each other.The annual mean correction factors according to Eq. ( 1) and their deviation from the combined correction factor of the two-year period are therefore shown in Table 3.
According to Table 3, the constant correction factor of the TEOM is very constant over the two-year period with a deviation of 7% to the single year factors.As mentioned, the SDS-heated-2 sensor suffered from malfunction during the first year.However, comparing the remaining correction factors of the sensors with heating, it can be noticed that they only have a maximum deviation of approx.15% from each other.The correction factors of the SDS-sensors without heating also agree very well with each other.The deviation of the correction factors for the PM2.5 measurements is higher than for the PM10 measurement.This is partly because only the TEOM, without a gravimetric measurement to compare it with, could be used as the only reference method and its correction factor could only be determined over a short period after the two-year measurement period considered here.However, for the PM10 and PM2.5 concentrations, a constant mean correction factor for all tested SDS011 sensors with heating unit may be applied.Therefore, the mean value for the two-year period of the yearly factors from the single heated SDS011 sensors is calculated (see Table 3).In this case the constant factor would be 2.51 ± 0.15 for PM10 and 2.59 ± 0.3 for PM2.5 concentrations.Whether this factor can be used as a generic factor for SDS011 sensors requires validation with a larger number of sensors.

Correlation of Sensors and Reference
Fig. 5 shows scatter plots of the sensor and TEOM data versus reference data, along with the regression coefficients R² as well as the slope and y-intercepts from linear regressions.Finally, the measured values of the respective three sensors with heating unit and four sensors without heating unit were combined by applying the calculated constant correction factor.The different x-and y-axis values in Fig. 5 should also be taken into account.
The TEOM, as equivalent reference method, expectedly showed the highest correlation with the filter measurement, however with R 2 = 0.6258, the correlation is still lower than expected.For this reason, the other correlation results for the hourly mean PM10 data and measurements  of PM2.5 concentrations measured by the sensors and plotted against TEOM as reference results should also be viewed with caution.Nevertheless, the y-intercept of the TEOM vs. filter reference regression is closer to 0 (1.097 µg m -3 ) and the slope is closer to 1 (1.045) compared to the SDS011 vs. filter reference regressions.The SDS011 sensors with heating unit, with an R 2 of 0.4694, show a much higher correlation with the reference measurement than the SDS011 sensors without heating unit with an R 2 of 0.08769.One reason for the low R 2 of the SDS011 sensors without heating are the days where the sensors drastically overestimated daily mean PM10 concentration of over 600 µg m -3 during humid days.This large overestimation of mass concentration is no longer as pronounced with the SDS sensors with heater, which significantly increases the correlation with the reference filter measurement.The overall scatter of the results from the sensors without heating unit is much more pronounced than of the ones with heating unit, resulting in significantly reduced R 2 values.This behaviour can be seen from the SDS011 sensors without heating in all the graphs shown in Fig. 5.The PM2.5,d values from the SDS011 sensors with dryer show with an R 2 of 0.4076 and a slope of the linear regression fit of 1.28 a similar picture as the PM10,d values with a slope of 1.248, although here the TEOM was used as a reference.Furthermore, in Fig. 5 it must be taken into account that the measured values of the SDS011 sensors were corrected with a common correction factor.Although the 7% deviation of the correction factors from the combined correction factor of the SDS011 sensors with heating unit (Table 3) is quite small, it also contributes to a higher scatter of the measured values shown in Fig. 5.However, looking at the individual sensors, the y-axis intercept of the regression line changes due to the constant correction factor.
Considering the y-axis intercept under the assumption that it can vary by up to 7% due to the correction factor, the deviation from the optimal result is marginal.It can further be noticed that the slope of the linear regression is greater than 1 for the daily averages for all sensors but below 1 for the hourly averages.During measurement, the TEOM may load the filter on the oscillating microbalance with small water droplets, resulting in elevated mass concentration readings.By heating the TEOM upstream, the previously separated water evaporates again, so that even negative mass concentrations can be output.When averaging the TEOM data, these negative readings are taken into account as both positive and negative measurement errors cancel each other out over time.Therefore, a negative 24 h mean concentration value is extremely rare.However, for the hourly means, negative values may occur.These were disregarded in the further comparison, resulting in an overestimation of the mass concentration in the remaining hourly TEOM averages.It can therefore be assumed that the TEOM overestimates the mass concentration in the hourly mean values, resulting in a slope of less than 1 in the linear regression with the SDS011 sensors.In comparison to the daily mean values, the hourly mean values of the PM10 and PM2.5 measurements are much more scattered.Because short-term fluctuations are partially balanced by averaging over 24 hours, the data are not as scattered as the hourly averages.Consequently, the R 2 values for these measurements are much lower at 0.2973 for PM10,h and 0.2658 for PM2.5,h.
Comparing the values measured with the SDS011 sensors with and without heating unit overall, a clear improvement in the measured values can be seen due to the aerosol drying upstream the sensor inlet.It also looks like a constant correction factor would be suitable for all SDS011 sensors with such an upstream aerosol drying.Although the constant correction factor contributes to a slight offset of the measured data compared to the reference, it has the advantage that the effort of using such a correction factor is minimal.The daily averages show a relatively good correlation with the reference measurement with R 2 = 0.4694, making it suitable for use within a dense monitoring network for the detection of fine dust sources.But if the individual sensors could be calibrated, this would lead to a more accurate measurement from the SDS011 sensors with heating unit within the measurement network.However, the readings are still significantly less accurate than type-approved methods, making it unsuitable for official particulate matter monitoring.Nonetheless, considering the price of about 100 € per complete measuring device including the aerosol dryer, they can be deemed acceptable, particularly given the strongly reduced effect of relative humidity.

SUMMARY AND CONCLUSIONS
The performance of four SDS011 low-cost PM sensors has been tested simultaneously in field measurements during a two-year period at a measurement location in Duisburg, Germany.The main objective of the study was to investigate the efficacy of a self-developed low-cost aerosol dryer used upstream of two of the sensors.The dryer heats the aerosol before entering a sensor in order to evaporate any particle water content.The other two sensor units were set up according to the recommendations by a German citizen science group (https://sensor.community/en/sensors/airrohr/).The aerosol dryers did not show any malfunctions during the two-year period of operation.The PM10 and PM2.5 concentrations, measured with the sensors, were averaged to obtain 1 h and 24 h mean values and compared to reference data from a TEOM and a low volume filter sampler.
The aerosol dryers effectively reduced the humidity of the aerosol and the measurements with the corresponding sensors were thus largely unaffected by the hygroscopic growth of particles due to atmospheric humidity variations.Only during cold winter days, the drying efficiency appears to have been insufficient.In contrast, the sensors without upstream aerosol drying produced significantly overestimated PM concentration during periods of high relative humidity.The results further show that the PM concentrations delivered by sensors at low relative humidity levels, e.g., because of the upstream drying, were too low and thus needed to be corrected.The 24 h mean values of the SDS011 sensors with heating units were on average by a factor of around 2.5 too low, whereas the measured values without heating were by a factor of up to 7 too high.The results show that an overall correction factor might be sufficient to produce data with reasonable accuracy without the need for regular and individual recalibration.For long-term measurements over one year, the simultaneous operation of two sensors at the same location appears to be a feasible way to control the sensors concerning possible malfunctions.In summary, SDS011 sensors were able to produce reasonable 24 h average PM10 and PM2.5 concentrations with the tested aerosol dryer over a period of one year.Furthermore, the heating unit tested has proven to be very robust and operate without malfunctions over the two-year period.

Fig. 1 .
Fig. 1.Measuring unit consisting of SDS011 sensor and aerosol dryer, shown without (top left) and with insulation (bottom left) and inside a waterproof case (right) (Nothhelfer, 2020).

Fig. 2 .
Fig. 2. Time series of the (a) temperature and relative humidity (data from weather station), (b) PM10 concentration and (c) PM2.5 concentration as 24 h averages; some SDS-3 and SDS-4 values exceed the scale at (b) and (c).

Fig. 3 .
Fig. 3. Monthly mean values of the (a) temperature and relative Humidity and correction factors FS for the different months during the measurement period for (b) PM10 and (c) PM2.5 (period of malfunction of SDS-heated-2 marked with bold orange curve).

Table 1 .
Coverage of the PM10, PM2.5, temperature (T) and relative humidity (r.h.) data from the different instruments over the two-year measurement period.

Table 2 .
Standard deviation of the monthly correction factors to the mean correction factor of year 1 and year 2 in [%].

Table 3 .
Annual and two-year correction factors for TEOM and the SDS011 sensors used.