Improving Urban Air Quality Measurements by a Diffusion Charger Based Electrical Particle Sensors – A Field Study in Beijing , China

ABSTRACTHigh aerosol loadings contribute significantly to the air quality problems of Asian megacities. To address this, monitoring data for aerosol mass and number that is spatially and temoprally of high resolution is needed, while the cost of obtaining such data remains high. Here, we present a field study in a polluted megacity, Beijing, using a diffusion-chargebased electrical aerosol sensor, the Pegasor PPS-M, which is a robust and comparatively low-cost instrument for the monitoring of both aerosol mass and number simultaneously. We present data over several months in the year 2014, and for varying aerosol size distributions, and analyze the performance against particle number and mass (volume) measured using a wide range particle sizer (WPS) and beta-attenuation-based PM2.5 observations. We show that using a single trap voltage, the PPS-M correlates well with particle mass, but not so well with particle number due to the variability in particle size distributions. However, the instrument response to number was improved by running the instrument with a variable trap voltage, and using the ratio of the different signals to gain information on the particle average volume. With this method, we were able to improve the correlation of the PPS-M; with the observed particle number from R = 0.14 to R = 0.72 for the measurement time period. Altogether, the PPS-M instrument displayed robustness and low maintenance requirements, and it showed good correlation with the other instruments in this study.


INTRODUCTION
Low air quality is a burning issue in many of the world's megacities, including for example Beijing, China (e.g., Chan and Yao, 2008).Concern for the health of the population of these cities (Liu et al., 2013) has led to increasing efforts to reduce particulate and gaseous air pollution in these cities.The complexity of the physics and chemistry of air pollution presents significant challenges to these efforts, and it is generally thought that more observation data of air pollution properties, especially aerosol mass and number concentrations, is urgently needed.
Aerosol concentration measurement is based on detecting a range of particle properties, such as number, mass, surface area, or volume (Kulkarni et al., 2011).For observation of atmospheric aerosol concentrations and the processes affecting them, instrumentation capable of real-time monitoring is desired.Such instruments are usually based on electrical or optical detection techniques.Optical instruments generally make use of light scattering or absorption by particles, while electrical instruments are usually based on first charging the particles, and the later measurement of the electrical current representing the charge carried by the particles.
Particle charging for charge-based techniques can be accomplished by unipolar diffusion chargers based on corona discharge (Intra and Tippayawong, 2009), and combined with a faraday cup electrometer, this produces a very simple aerosol detector.Such detectors have been used for various purposes, such as diesel engine emission measurements (Ntziachristos et al., 2004) and lung-deposited surface area monitoring (Fissan et al., 2007).Small, even handheld versions of such instruments exist (Marra et al., 2009;Fierz et al., 2011;Lee et al., 2011).
The above-mentioned electrical instruments rely on the collection of particles after charging.For observations in conditions with high aerosol loadings, such as very polluted megacities, collection-based measurement may be problematic due to the accumulation of aerosol in the instrument.However, charge-based observations are possible also without particle collection; this was demonstrated by Lehtimäki (1983) with an instrument based on measuring the charge escaping the charger with the charged particles.The same principle has been developed and applied by Rostedt et al. (2009b), andrecently, Fierz et al. (2014).
In this paper, we use the Pegasor PPS-M (Pegasor Oy, Tampere, Finland) sensor, which is based on the escaping charge principle, which was presented by Lanki et al. (2011).Ntziachristos et al. (2011Ntziachristos et al. ( , 2013) ) have successfully applied the sensor previously to diesel exhaust measurements and Järvinen et al. (2014) showed that the PPS-M can be used to observe ambient urban aerosol in relatively clean, Nordic conditions.Here, we present data from a measurement period in the highly-polluted megacity of Beijing, China, and utilize that in order to optimize the use of sensor trap function.The aim of this study was to deploy the PPS-M instrument in a highly-polluted urban setting over a longer time period to evaluate the instruments response in varying aerosol loadings.Because the response of the PPS-M depends on the size distribution of the ambient aerosol, we compared the PPS-M data with particle size distribution data.In a recent study, Amanatidis et al. (2016) have demonstrated that using a dual PPS-M setup, with each instrument operating at a different trap voltage, can be used to gain information on particle size.
A key parameter of atmospheric aerosol is the condensation sink (CS, see e.g., Dal Maso et al., 2002).It represents the aerosol's ability for irreversible vapor uptake via condensation (Riipinen et al., 2012), but also gives the timescale of small particle scavenging (Kerminen and Kulmala, 2002;Lehtinen et al., 2007).In addition, at least for size distribution observed at a background site, the condensation sink is proportional to the ion sink of an aerosol and measurements at the same background site have shown that the current of an instrument using unipolar charging is proportional to the condensation sink (Kuuluvainen et al., 2010).As the PPS-M also uses unipolar corona charging to charge particles, we aimed to investigate whether a similar proportionality could be found for the PPS-M and for heavily polluted air.The condensation sink can be computed from the size distribution according to e.g., Dal Maso et al. (2002).
Here, we investigate the differences in the signal of a single PPS-M with varying trap voltages, and use the data to gain similar information from a single PPS-M, to explore the possibilities of using this information to develop sensor type air quality measurement to better correlate with ambient air particle number and mass measurement.Based on these aims, this study will a) describe PPS-M measurements and the analysis of the data in the polluted Beijing megacity over a period of 153 days b) present a general overview of the types of aerosol size distributions and their variation observed at the site of observation; c) analyze and develop the sensor type measurement for ambient aerosol number, mass, and active surface area concentrations.

METHODS
All the data used in this study were collected in the campus of The Chinese Research Academy of Environmental Sciences (CRAES), which is located near a residential area in the northern part of Beijing (116°24'E, 40°02'N) outside the fifth ring road of the city (Wang et al., 2009).The nearest large road is situated ca. 100 m south of the measurement station.The instruments were set up in a three-room observational station located on the rooftop of a three-floor building in the Academy (Wang et al., 2010;Gao et al., 2012).
We used two PPS-M sensors in the measurements.The first PPS-M instrument was installed at the CRAES field observation site in February 2014, the second in September 2014.The sample inlet was situated on the cabin roof, shielded from rain but otherwise without any size cutoff.The instrument exhaust was led out thought the cabin wall to ensure pressure balance, but the exhaust was situated away from the inlet to avoid disturbing the ambient charge balance near the inlet.To dry the sample the PPS-M inlet was heated up to a temperature that was 40°C above the ambient temperature, ensuring that ambient humidity does not condense inside the instrument.Background measurements (PPS-M signal at zero particle loading) were performed regularly using a HEPA filter applied to the PPS-M inlet, and if present, the background signal was removed during data analysis.A droplet separator was added to the pressurized air line as a caution to remove possible water droplets in the line generated by the pressurized air pump.One should note that The PPS-M data was collected in two ways: via an Ethernet controlling box, which transmitted collected data to a server connected to the internet, and also via a USB-connected PC situated in the measurement hut.The USB-connected data was used in this study.The PPS-M instruments were installed at a measurement station housing also other aerosol instrumentation, such as the Wide-Range Particle Sizer (WPS) and the second reference instrument deployed at CRAES, a beta-attenuation -based instrument that measured the mass concentration of PM 2.5 particulate matter (Gao et al., 2007;Wang et al., 2014).An aerosol Wide-range Particle Spectrometer (WPSTM model 1000XP, MSP Corporation, USA) (see e.g., Gao et al., 2012) combines the principles of differential mobility analysis (DMA), condensation particle counting (CPC) and laser light scattering (LPS), was used to measure particle counts in the range of 0.01 to 10 µm.The measurement setup is shown in Fig. 1.

Single PPS-M Observations
The PPS-M measurements at the CRAES measurement site were started on February 22 nd 2014 and the first observation period lasted until July 25 th .The maintenance performed on the instrument was minimal involving routine checks on the instrument and data saving from the computer, and approximately every four weeks the instrument was disassembled to change the corona needle inside the instrument.
The PPS-M was set up to record data every second; however, to compare with the WPS data, the PPS data was averaged to a 3-minute average value corresponding to the WPS measurement time.The relationships between data observed by different instruments were studied by computing Pearson correlation coefficients (R) between the different datasets.In cases of varying time resolution of the data, the data with higher sampling rate was averaged over the sampling time of the instrument with the lower sampling rate.

Basis of PPS-M Trap Voltage Scanning
In their study, Amanatidis et al. (2016) show that assuming a log-normal and reasonably well-behaved size distribution (e.g., assuming aerosol from a combustion source), the total aerosol mass-to-number ratio can be expressed as an analytical function of the size distribution parameters, including the geometric mean diameter, which in turn can be estimated from the ratio (R Ut ) of two PPS-M instruments running with different trap voltages.However, studying the size distribution data from the WPS, we found that for a significant amount of time, the size distribution was not unimodal, and thus many of the assumptions did not hold.Therefore, we adopted an empirically grounded method in which we obtained the ratio of the PPS-M current obtained in subsequent measurements using different PPS-M trap voltages, and used the R Ut obtained thus as a parameter to compare with the observed average volume F = V tot /N tot , where V tot and N tot are the total particle volume and number, respectively, measured by the WPS.By comparing these observations, we aimed to find an empirical correlation that could be used to estimate the average volume of particles for a variable aerosol.The details of the developed empirical correlation are discussed in the Results section.data coverage was 21% of the whole period, comprising of 15300 individual size distribution and PPS-M observations.Data outages were ca 10% more frequent during morning hours (9-10 a.m.) due to data saving and upkeep occurring during these times.Overall, no statistical bias toward any time of day due to overrepresentation in the data should, however, be observable.Minimal bias (3.5% higher than expected) towards weekend data could be observed, presumably for the same reasons.Weather conditions or pollution levels were never causes for instrument outages.All in all, we assume that our data was representative for Beijing megacity air during the measurement period.

RESULTS AND DISCUSSION
An overview of the frequency distributions for key parameters in the WPS and PPS-M data are shown in Fig. 3. Particle number concentration in the measurement period was observed to be between a few thousand and up to 10 5 particles per cubic centimeter, with a mean concentration of 31800 # cm -3 and a median of 21300 # cm -3 .The mass concentration, calculated using an average density of 1.6 g cm -3 , varied between a few tens up to several hundred micrograms per cm 3 , in line with the highly polluted environment.The mean mass concentration was 62 µg m -3 , and the median 43 µg m -3 .The density estimate is based on approximate the average measured densities of Beijing air by Liu et al. (2015).They reported density with high variability, and in principle one could possibly improve on the density estimate by using a regression function to our data; however, this was considered outside the focus of this study.
PPS-M current values showed a more varied frequency distribution, as can be expected due to the size-dependent response of the instrument.The mean current value observed in the first observation period was 0.37 pA with a median of 0.35 pA.The frequency distribution was less skewed than size distribution parameter distributions, resembling a normal distribution.The standard deviation of the observations was 0.22 pA.

Single PPS-M Comparison with Beta-Attenuation -Based PM 2.5 Observations
The hourly values of PM 2.5 were compared to the arithmetic mean of PPS-M observations obtained from the one-second data, and the result is shown in Fig 4 .One can observe the changes in PM 2.5 as the season transitions from winter to summer; winter PM 2.5 is seen to be very high, while spring and summer values mostly stayed below 150 µg m -3 .The PPS-M values generally follow the mass observations very well, and the Pearson correlation coefficient R = 0.82 (R 2 = 0.66), which can be considered very good for two independent instruments measuring urban air; for reference, shorter and hourly-averaged measurements at roadside and residential areas comparing PM 2.5 and PPS-M data resulted in Pearson correlations ranging from 0.3 to 0.74 (Järvinen et al., 2014).It should be kept in mind that in our measurements, the PPS-M was used with 400 V trap voltage.

PPS-M Comparison with WPS
Over the first measurement period, PPS-M values generally followed WPS total particle volume readings well.In Fig. 5(A) we show a scatterplot for the observations, and the correlation for the WPS-measured mass and the PPS-M current is high, with R = 0.86 (R 2 = 0.70).For the number concentration, the situation is different; the correlation between the total number and the PPS-M current was low.This reflects the variability of the Beijing aerosol in terms of size distribution and total number concentration; the correlation between total WPS-measured volume and total number concentration was low (R = 0.13, R 2 = 0.02) This differs from the situation that was observed e.g., in Järvinen et al. (2014), where the measurement environment was more homogeneous for each situation.From the scatterplot (Fig. 5(B)) one can clearly observe that for some situations, a correlation might exist, but the relation did not stay consistently similar.Thus, Fig. 5(B) clearly shows the need for improvements in sensor operation in environments were the monitored aerosol, especially its particle size distribution, changes significantly.This could be achieved by changing trap voltage of the sensor described later in this paper.
To explore the relation between the number concentration and the PPS-M reading in more detail, we studied the correlation between the condensation sink and the PPS-M signal in addition with the number concentration (Fig. 6).Theoretically, one would expect high correspondence between CS and the PPS-M signal, as the charging efficiency of aerosol particles is proportional (and also phenomenologically similar) to the sink of non-volatile vapor to particles.However, as can be seen from Fig. 6, the response of the PPS-M to condensation sink is not linear; rather, we can observe that the response changes when the total particle number changes, with higher particle numbers    leading to a relative reduction of the PPS-M current response.This can be explained by considering the function of the ion trap/mobility analyzer part of the instrument: the ion trap penetration at 400 V is 75% at 40 nm (Rostedt et al., 2014), and much lower for smaller sizes.In situations where number concentrations are high, the size distribution is often dominated by small particles, and therefore the PPS-M responds less efficiently to a rise in the CS.This observation gave cause to use the particle size-dependent penetration function of the PPS-M mobility analyser (ion trap) to gain information on the volume-to-number ratio of the aerosol.
One should note, that the heated inlet used results in drying the air while warming, so if some of the aerosol is hygroscopic, their size might be reduced.However, drying the aerosol sample before measurement is quite often used in long-term measurements due to the operational benefits; it also reduces the uncertainty that would be caused by the rather arbitrary changes in RH when the air would be brought into the instrument from outside to room temperature.

Scanning the Trap Voltage of the PPS-M
The analysis for the first period of observations showed that satisfactory response for particle mass could be obtained from the PPS-M at 400 V, but to gain good information on the particle number, additional information on the mass-tonumber ratio is needed.This could in principle be achieved by variation of the trap voltage (mobility analyser, Rostedt et al., 2014) of the PPS-M.
In general, Rostedt et al. (2014) showed that varying the trap voltage affects the penetration of the mobility analyzer part of the instrument in a size-dependent manner.This, combined with the prevailing atmospheric property that high number concentrations in urban areas are related to smaller particles (see e.g., Mönkkönen et al., 2004;Pirjola et al., 2012) gave reason to expect that differential comparisons of the PPS-M signal at varying trap voltages should add to information on the number-to-volume ratio.To test the applicability in the field of the variation of the trap voltage, the instrument controlling unit was modified and set to scan three different trap voltages: 50 V, 400 V and 1000 V.The normal setting of the trap is 400 V. Duration of one scanning period was one minute.
During the trap scanning observations, two instruments were operated side by side at the CRAES observation site.The instruments were set up identically, with the only difference being that the inlet flows of the two instruments were 2.4 and 2.7 L min -1 , respectively.The instrument signals were highly correlated (see Fig. 7(A); Pearson R = 0.999) over the whole observation period; in addition, also the signal ratios, defined as R Ut = I Ut=50V /I Ut=1000V were close to each other (see Fig. 7(B); Pearson R = 0.96).The higher deviations of R Ut are due to the variation in the flow rates, which affect the shape of the trap penetration curve, and therefore also the signal ratio in a non-linear manner.
We computed R Ut for each PPS-M measurement cycle, and then averaged the values over the time of the WPS spectrum measurement.From the WPS spectrum, the total volume was obtained by assuming spherical particles and computing the total volume from size distribution data.As a result, we obtained the average particle volume (computed as V tot /N tot from WPS data) as a function of R Ut (Fig. 8).As can be observed from the figure, low particle average volume (corresponding to small particles, and usually to high number concentration) corresponds to a higher R Ut , in line with the assumption that smaller particles get efficiently removed by the high trap voltage, while they mostly penetrate at the low trap voltage.To this empirical correlation, we fitted a function of the form where a = 3.33 × 10 21 and 1.84 × 10 21 and b = 5.34 × 10 21 and 3.16 × 10 21 for sensors 4601 and 4608, respectively.The difference in the fitting parameters can be explained by the differences in the inlet flow rates and the resulting changes in the penetration efficiency.Based on the observations, the flow-corrected number-to-volume ratio can be semi-empirically approximated with   where f is the inlet flow rate, f 0 is a reference flow rate (in this case the inlet flow rate of sensor 4601, 2.7 L min -1 ), and c = 4.2 is a sensitivity parameter describing the signal sensitivity to the inlet flow rate.a 0 and b 0 are the reference parameters for sensor 4601.Using a zero-forced regression for the WPS volume as a function of the PPS-M signal with U t = 400 V (Fig. 9), and combining this regression with the function F, we can get an expression for particle number as a function of only I Ut=400V and R Ut : (3) here m = (f 0 /f) × 5.68 × 10 -17 is the fitting coefficient for the WPS volume signal.An example of comparing WPS number   with the PPS-M derived number using Eq: (3) is given in Fig. 10(A).One can see that the number model performs well even for peaks with very high number concentrations.Over the whole period, the model performance can be considered good, with a coefficient of determination R 2 = 0.52 (see Fig. 10(B)).
The trap scanning method requires longer time to obtain a single obvservation, reducing the time resolution of the PPS-M instrument from up to 10 Hz to approximately 1 minute per observation.However, in return one can obtain key new information on the dominant size of particles, which is a key factor in determining the origin and fate of airborne particles.Compared to the dual-sensor setup of Amanatidis et al. (2016), the setup also suffers from a lower time resolution; however, it lacks the drawbacks of having to account for potential differences in the flow rates in the instruments, as the differences in trap voltage are obtained in a single instrument.This also makes the oneinstrument approach the more economical solution.As the diffusion charging-based sensors such as the PPS-M are fairly inexpensive when compared to other instruments, operate well also on very polluted environments, and require comparatively low maintenance, this potentially improves their attractiveness as components of urban air quality monitoring networks in highly polluted areas such as Asian megacities and industiral centres.

CONCLUSIONS
In this study, we have deployed diffusion charging instruments (Pegasor PPS-M) based on the escaping charge principle at a highly polluted Chinese megacity observation site, and compared the sensor signals with PM 2.5 observations, aerosol size distribution observations, and inter-compared two PPS-M sensor with each others.The aerosol at the observation site was variable: at times, the number size distribution was dominated by very small particles (ca 20 nm in size), while at other times, larger particles with a mean diameter of >100 nm were abundant.We found that the PPS-M sensors were robust and required low maintenance, and performed well over extended periods of time.The raw current signal of the PPS-M instrument using trap voltage V trap showed high correlation with the beta-attenuationderived PM 2.5 observations and the wide-range particle sizer (WPS) -based particle volume and mass observations.However, the raw current did not correlate well with the total number concentration observed by the WPS, due to the high variability in the size distribution.
In an attempt to improve the correlation with the PPS-M signal, we set up two PPS-M instruments in Beijing with a measurement routine that used three different trap (also called the mobility analyzer) voltages.We found that the ratio of the signals obtained with the maximum and minimum trap voltages were proportional to the average particle volumeobserved by the PPS-M, and derived the dependency function.Using only the signal of the PPS-M, using the signal ratio at different trap voltages, we could significaltly improve the correlation between the PPS-M and WPS number concentration observations, from (R = 0.14 to R = 0.72 for PPS-M without trap scanning and the inversion involving trap scanning information, respectively).Overall, the PPS-M was found to be applicable as a megacity aerosol sensor, and the variable trap voltage method, while reducing the time resolution of the instrument, produced information on the average size of particles.

Fig. 1 .
Fig. 1.Experimental setup for the single PPS-M measurement at the CRAES station.

Fig. 2 .
Fig. 2. Time series of the PPS-M sensor output (sensor current in pA) in Beijing in 2104.

Fig. 3 .
Fig. 3. Frequency distribution histograms for four key parameters of the aerosol observed at CRAES in 2014: (i) particle total number, (ii) particle total mass, (iii) the aerosol condensation sink and (iv) the sensor current signal observed by the PPS-M.Parameters (i)-(iii) were calcultaed from WPS data.

Fig. 4 .
Fig. 4. Comparison of the hourly-averaged PPS-M signal to the hourly beta-attenuation-based PM 2.5 signal at the CRAES measurement station in 2014.The top panel shows time series of both instrument readings, while the bottom panel shows a scatter plot of the hourly value with the Pearson correlation coefficient.

Fig. 5 .
Fig. 5. Correlation scatter plots of the PPS-M signal compared to total mass (left panel, calculated from the size distribution assuming spherical particles and constant density) and particle total number (right panel).

Fig. 6 .
Fig. 6.A Scatter plot displaying the PPS-M current signal, the computed condensation sink and the total particle number (color coded) at the CRAES station during the single-PPS observations in 2014.

Fig. 7 .
Fig. 7. Scatter plots comparing the two sensors (No:s 4601 and 4608) installed later in the year 2014 at the CRAES observation station.The top panel shows the intercomparison of the simultaneus observations of the two sensors, while the bottom panel shows the intercomparison of the ratio of the signals at trap voltages V trap = 50 V and V trap = 1000 V for both sensors.It should be noted that for the latter intercomparison, the two instruments were not simultaneously at the same trap voltage setting, which causes a temporal offset of the two values of 20-40 s.

Fig. 8 .
Fig. 8. Mean particle volume (V tot /N tot ) observed by the WPS instrument as a function of the signal ratio at V trap = 50 V and V trap = 1000 V for the PPS-M instrument.Observed values are given as points, while the black line shows the fitting function given in Eq. (1).

Fig. 9 .
Fig. 9. Correlation scatter plot of the PPS-M signal at trap voltage V trap = 400 V and the total volume observed by the WPS instrument during the PPS-M trap voltage scanning period in 2014 for sensor 4608.

Fig. 10 .
Fig. 10.Comparison of the time series for the particle number observations of the WPS and the value obtained by applying Eq. (3) to the signal obtained from the PPS-M for the trap scanning experiment period in 2014.The top panel shows the time series of the observed WPS number signal, while the bottom panel shows the scatter plot with the values thor the corresponding observation times.The PPS-M inversion values were averaged for the time period (3 minutes) of the WPS scanning.