Using low-cost sensors to monitor indoor, outdoor, and personal ozone concentrations in Beijing, China

Abstract

High concentrations of ground-level ozone (O₃) have been measured outdoors across China but there are limited measurements of O₃ in microenvironments, including in homes, and for personal exposure. This highlights the need for cheaper methods to accurately make these measurements and to better capture fine-scale spatial variability in O₃ across cities. With this in mind, we conducted a pilot study at six homes in Beijing, China, over 12 days to evaluate the use of portable, low-cost, time-resolved monitors for measuring O₃ indoors and outdoors. We also assessed personal exposure for one adult in each home for two 48 hour periods using backpack-mounted monitors. Prior to and following sampling we collocated all monitors with a reference analyzer; we used data from these colocations to generate linear calibrations which we applied to all monitor data. Calibration slopes did not change significantly over the study although some intercepts differed. The average limit of detection (LOD) was 7.0 ppb, average root mean square error was 16.7 ppb, mean absolute error was 13.3 ppb and normalized root mean square error was 33%. Performance varied substantially between sensors, underscoring the importance of monitor-specific calibrations and determinations of measurement error. Outdoor concentrations varied spatially, with home-specific peak hourly averages of 32–165 ppb; indoor concentrations ranged from below the LOD to 15 ppb. Hour-averaged personal exposure was generally higher than O₃ indoors, and at times exceeded ambient O₃ indicating contributions to personal exposure from ambient sources of O₃ away from the home. This work illustrates the feasibility of using these monitors to characterize distributions of O₃ spatially and temporally when differences in concentrations are large, and outlines considerations for using these monitors to measure personal exposure.

---

Environmental significance

Ground-level ozone (O₃), which affects both human and environmental health, has become one of the main harmful air pollutants in China. In spite of this, limited data exists about personal exposure to O₃ and about concentrations of O₃ in microenvironments across a region. In this study, we illustrate the feasibility of using low-cost sensors to address these knowledge gaps and describe considerations when calibrating the sensors to improve data accuracy. We highlight the large gradients in concentrations and exposures to O₃ that can be measured in a relatively polluted urban area (here, Beijing) using these small, portable, real-time sensors and show that if homes are well-sealed, the concentrations of O₃ indoors are often quite low.

1. Introduction

Ground-level ozone (O₃) pollution is a global concern, and O₃ has become one of the most problematic air pollutants in China with average daily concentrations reaching as high as 242 μg m⁻³ (124 ppb) in Beijing.¹ The Chinese National air quality standards are 100 μg m⁻³ (51 ppb) for Class 1 O₃ concentrations and 160 μg m⁻³ (82 ppb) for Class 2 O₃ concentrations, measured as daily 8 hour maximum averages.² Class 1 standards apply to areas in need of special protection, i.e., national parks, while Class 2 standards apply to urban and industrial environments.¹ Many other cities in China including Shanghai, Nanjing, Xi 'an, Jinan, and Zhengzhou also have high concentrations of ground-level O₃, with concentrations highest from July to September.³ Factors such as high temperature, strong sunlight, and emissions of precursor pollutants (i.e. NO_x and specific volatile organic compounds) can contribute to elevated concentrations of O₃ throughout the year with values typically highest in summer due to the increased presence of sunlight (UV irradiation). Monitoring ground-level O₃ is an important first step in understanding its spatial and regional patterns and concentrations, which are necessary in assessing environmental and health effects.

Current approaches to measuring O₃ concentrations are limited, in part due to the instrumentation available.⁴ Conventional air pollution monitoring systems typically employ instruments with relatively complicated methods of measurement and use additional tools or equipment including temperature and relative humidity controllers, and air filters or sensors for particulate matter and gases. They also have built-in calibrators, aiming to guarantee the accuracy and performance of the instruments. Consequently, these instruments tend to be large, expensive, and work best with a constant supply of stable power.⁵ These factors limit their use to environments with a reliable power supply and where theft or vandalism of expensive equipment are not a major concern and make it challenging to use them in smaller spaces including homes or to transport instruments readily between locations. These limitations further hinder the development of systematic networks of time-resolved monitoring instruments, resulting in the inefficiency of the monitoring of strong gradients in pollutant concentrations both spatially and temporally.⁴ Understanding the impacts of air pollutants on human health and the environment can be challenging without these data. Small, low-cost (e.g., costing a few hundred dollars) sensors which can collect measurements integrated over short intervals of time (e.g., minutes or hours) provide a means to fill these gaps in the data.⁵ While the potential for such devices is well understood, low-cost O₃ sensors have not been used as widely as might be expected due, in part, to the challenges of obtaining precise and accurate measurements of specific gaseous pollutants using low-cost sensors, especially over a long period of time, as compared to using conventional monitoring instruments.^5,6 Detailed published measures of performance of low-cost gas sensors are relatively sparse and each low-cost sensor must be thoroughly evaluated to understand its limitations and to determine the best methods of calibration to maximize data quality.^7–10 Moreover, networks of time-resolved monitoring instruments or systems are not being widely applied in China.

This study was a pilot preceding an extensive field campaign in Shanghai; as such, an overarching goal of the research was to evaluate the performance of a particular type of low-cost O₃ monitor and to determine how well the methods of deployment and calibration work in these highly polluted environments in urban China where there are large temporal–spatial and indoor/outdoor variabilities in concentrations of O₃. The specific aims of the pilot study were to (a) test portable, low-cost, time-resolved monitors for measuring concentrations of O₃ indoors and outdoors at homes in Beijing, (b) to measure personal exposure to O₃ and (c) to evaluate the performance of these monitors for these purposes.

2. Methods and equipment

2.1 Study location and participant recruitment

This study was conducted in Beijing, China (40° N, 116° E, 43 m asl.). Beijing is the world's third most populous city proper, with a population of over 17 million and an area of 16 800 km² (38% flatlands and 62% mountain areas). Beijing is surrounded by high mountains to the north, southwest, and west.¹¹ This geography makes Beijing vulnerable to air pollution because easterly or southerly winds can transport large amounts of pollution to Beijing where it accumulates.¹¹ From 2010–2014, the average concentration of O₃-8 h in urban Beijing increased at a rate of 1.8 μg m⁻³ (∼0.9 ppb) per year, with concentrations in 2010 of approximately 80 μg m⁻³ (∼40 ppb).¹² The combined effects of urbanization and pollution caused by combustion of fossil fuels has resulted in Beijing becoming one of the most polluted cities in the world with several intense air pollution events occurring each year.¹³

We included six homes from across Beijing in this pilot study (Fig. 1). Homes were selected through collaborators at Tsinghua University. As participants were selected from university contacts, we do not expect them to reflect the general population; however, as homes were located across Beijing, we expect them to illustrate a range of pollutant concentrations. Home 1 was located in a busy commercial area at the third-ring road in Beijing. Homes 2 and 3 were on the Tsinghua University campus in northwestern Beijing. Home 4 was in western Beijing, and homes 5 and 6 were in east Beijing. These latter three homes were in residential apartments and neighborhoods. During personal sampling, participants engaged in their usual routine, with some time spent at work, some time spent at home, and some time spent commuting. Participants who lived on campus also worked on campus.

Fig. 1 Sampling locations in Beijing, with regulatory monitoring stations indicated with a black dot.

Prior to sampling in the homes and on individuals, all monitors were collocated with a stationary reference analyzer (Thermo Electron Corporation UV Photometric O₃ Analyzer, Model 49i) at a base station at the College of Computer Science, Peking University (PKU) in order to ensure monitors were working properly and to derive calibrations for the monitors. The reference analyzer logged minute averages of O₃ in addition to temperature and relative humidity. All instruments were located on the roof of the building – five floors above ground level – and the site had relatively unobstructed airflow.

2.2 Description of low-cost monitors

We used indoor/outdoor and personal exposure monitors to measure time-resolved concentrations of pollutants (Fig. 2A). Monitoring packages, which cost approximately 350 USD, included 2-way (i.e., comprising oxidant and NO₂ sensors) or 3-way (i.e., comprising oxidant, NO₂ and NO sensors) Alphasense 4-Electrode gas sensors (OX-A4 and NO₂-A4 series), a Plantower PMS3003 particulate matter (PM_2.5) sensor, in some cases a Cozir CO₂ sensor, an SD card, LED indicators, a Sensirion SHT15 temperature and relative humidity (RH) sensor, a clock, a voltage regulator, a barrel jack for wall power supply, and a Teensy microcontroller. The Alphasense gas sensors were connected to Alphasense analogue front end (AFE) support circuits and these in turn, along with all other components, were connected to a printed circuit board (PCB) designed specifically for these monitors. Diodes, resistors and transformers on the PCB were used to convert the AC power supplied via wall power to DC power required by the sensors. Sensors were located flush with the edge of the monitoring box so that they were directly exposed to ambient air – this minimizes the possibility of O₃ reacting with the inside surfaces of the monitor before it reaches the sensor. When using these monitors, minute averages of gas phase pollutant concentrations are recorded on an SD card connected to the Teensy microcontroller, and data can subsequently be downloaded by inserting the SD card into a computer. An LED indicates the status of the monitor (off, logging data, not logging data, error). The barrel jack is used to connect the monitor to a portable battery or to wall power. The indoor/outdoor and personal monitors all contain the same components, although the external packaging differs and thus the components are arranged in a slightly different configuration in the packages. Fig. 2B and C show a personal monitor which can be attached to a backpack to evaluate time-resolved personal exposure to air pollutants. Although additional pollutants were measured using these monitors, in this paper we focus only on the O₃ data; PM_2.5 data are discussed in a separate publication.¹⁴

Fig. 2 (A) A monitor used indoors and outdoors containing 3-way Alphasense gas sensors for measuring O₃, NO₂, and NO; (B) a monitor used for measuring personal exposure containing 2-way Alphasense gas sensors for monitoring O₃ and NO₂; (C) a personal exposure monitor hanging from a backpack. *Only a subset of packages includes NO and CO₂ sensors.

Alphasense gas sensors operate in the amperometric mode, generating a current that is linearly proportional to the fractional volume of a specific gas. The sensors comprise four electrodes – a working electrode, a reference electrode, a counter electrode, and an auxiliary electrode. The working electrode is designed to optimize the oxidation or reduction of the gas to be measured, while the other three electrodes are used to balance the chemical reaction of the working electrode and provide a stable environment such as equivalent current and stable potential. The sensor's analog output ranges from 0–3 volts and is converted to ppb based on sensor-specific factors – total working electrode zero (mV), total auxiliary electrode zero (mV), and sensitivity to O₃ and NO₂ (mV ppb⁻¹) – provided by Alphasense. In these monitors, the conversion from sensor signal to oxidant and NO₂ concentration (ppb) is done using these factors via Arduino code uploaded to the Teensy microcontroller. As is the case for many electrochemical sensors, these sensors experience cross-sensitivity to multiple pollutants. The Alphasense platform thus uses an oxidant sensor – responsive to NO₂ and O₃ – and an NO₂-specific sensor in tandem to measure concentrations of O₃; the NO₂ sensor filters out O₃ using a manganese dioxide filter to convert O₃ to O₂, and the concentration of O₃ can be determined by subtracting the response of the NO₂ sensor from the response of the oxidant sensor. Previous work has shown these sensors detect O₃, NO₂, and NO with high sensitivity, stability, and reliability for long-term repeatability,^6,15 although one study found that concentrations of O₃ tend to be underestimated in mixtures of NO₂ and O₃.¹⁶ While O₃ concentrations are logged to the SD card, these monitor-derived concentrations should still be calibrated through colocation and comparison with a reference analyzer, as detailed in Section 2.3.1.

2.3 Study design and timeline

2.3.1 Monitor colocation, calibration, and determination of limits of detection. To calibrate the monitors, we first set up all devices alongside the reference analyzer and left them to run for 48 hours on the top of the PKU building described previously from July 27^th, 2016 to July 29^th, 2016. We repeated this colocation process after home/personal monitoring, from August 15^th, 2016 to August 16^th, 2016. We generated best-fit linear regressions between quality-controlled monitor data and data from the reference analyzer; this was done separately for the two colocation periods to determine if the calibrations for individual monitors changed over time. The regression constants, which we ultimately derived from colocation data from both periods combined (discussed in greater detail in Section 3), were subsequently used to calibrate monitor data collected in the homes and to calibrate personal exposure data. We also used these data collected during colocation periods at PKU to determine the agreement between the data from low-cost monitors and from the reference analyzer. To further evaluate the stability of the calibrations over time and to better understand the calibration requirements for these monitors (e.g., adequate duration of monitor/reference analyzer colocation) we left two of the low-cost monitors on the roof of the PKU building for the duration of the field campaign.

A limit of detection (LOD) was calculated for each monitor using data from periods when monitors were colocated at PKU and the reference analyzer recorded O₃ concentrations less than or equal to 3 ppb. The standard deviation (σ) of the readings of the monitors during these periods of low O₃ concentration was calculated for each monitor separately, and monitor-specific LODs were calculated as 3σ.¹⁷ Although we might expect concentrations of O₃ in the homes to be low due to rapid reactions with surfaces, they are likely not consistently zero; we believe that excluding values below the LOD altogether would bias the data. As such, LODs were applied to hour-averaged data collected during home and personal exposure monitoring and values below the LOD were replaced with 3/4 the LOD.

2.3.2 Collecting stationary measurements at homes. Following the initial colocation of all monitors at PKU, we measured air quality in six homes for approximately 12 days from August 1^st, 2016 (some homes started as early as July 30^th, 2016) to August 12^th, 2016. These O₃ measurements were collected as part of a larger effort to assess the effectiveness of portable air filtration devices on levels of particulate matter. As such, an Atmosphere® Air Purifier was installed in one bedroom of each home. During a ∼6 day period (“true filtration”) the purifier contained all filters (a coarse pre-filter, a replaceable HEPA [high efficiency particulate air] filter, and a replaceable carbon filter) and during the other ∼6 day period (“sham filtration”) it contained only the pre-filter. The order of filtration intervention varied randomly by household, and the interventions (true, sham) were carried out back-to-back in each home. We installed indoor and outdoor monitors inside and outside each home and let them run for the entire 12 days, with the indoor monitors located in the room with the air purifier. We asked participants to keep the doors and windows to the bedroom closed during sampling. Monitors were connected either to a wall outlet or a power strip, or in cases where this was not a viable option they were powered by battery. In most cases the indoor monitor was in the breathing zone between 1 and 1.5 meters above the floor. We mounted monitors via zip ties or placed them on a shelf or other surface, and we ensured the inlet and exhaust were unobstructed. After setting the monitors up we let them run for several minutes to guarantee that they were functioning, and that the data were being logged to the SD card.

2.3.3 Conducting personal exposure monitoring. In addition to the stationary monitoring, we asked one adult from each home to carry a small backpack to monitor personal exposure to O₃ for one 48 hour period during true filtration and one 48 hour period during sham filtration. Personal monitors were always powered by battery using batteries that can run for approximately three days when used with this load (Xiaomi, 20000mAH Powerbank). The backpack also contained a low-cost monitor and a small pump for collecting filter samples.¹⁸ We asked that participants wear the backpack at all times except when sleeping or bathing/showering. While sleeping, we asked that the pack be placed in the bedroom on a side table or shelf at about the same height above the floor as the bed. While bathing, the device could be kept in the bathroom but could not get wet.

2.4 Data processing

Minute averages of data (temperature, relative humidity, concentrations of O₃ and NO₂) were logged during this study. Most studies using these Alphasense O₃ sensors have employed these sensors to collect stationary ambient measurements or in well-controlled chamber studies, and to our knowledge we are one of the first to use these sensors to measure personal exposure. In a very recent study, these sensors were deployed in both a stationary and mobile configuration with minute-averages of data analyzed. In the aforementioned study, agreement with reference analyzer data was worse with the mobile configuration illustrating some of the challenges of collecting mobile data.¹⁹ In the current study, raw, one-minute data from the low-cost monitors and the reference analyzer were hour-averaged as one-minute data for both personal exposure (i.e. mobile measurements) and indoor/outdoor (static measurements) were relatively noisy and time-activity data were collected by hour. The decision to hour-average was made to balance the need for higher time-resolution with the need to ensure data robustness.

These Alphasense electrochemical sensors require time to stabilize upon startup, and outliers representing erroneous values occur frequently at the very beginning of sampling. To account for this warm-up period, we removed the first two hours of data after monitor start up. In some instances, the sensor became dislodged from its connection to the rest of the monitor resulting in very high and variable values (e.g., values several hundred ppb higher than those measured from all other sensors). These erroneous values were easily identifiable when data from numerous monitors were plotted simultaneously over time, and such data were removed prior to further analysis. Lastly, from the remaining data we removed hour-averaged data collected during the colocation periods that were greater than the third quartile (Q3) plus 1.5 times the interquartile range (IQR) or less than the first quartile (Q1) minus 1.5 × IQR of each single set of data (reference analyzer or monitor). This process preceded the determination of calibration constants for the monitors.

3. Results

3.1 Field evaluation of O₃ monitors

3.1.1 Data processing, calibration and monitor performance. During the colocations, monitors used indoors at home 6 and for personal exposure at homes 3 and 4 showed no, or few, obvious trends when compared with data from the reference analyzer. This may indicate that the sensors were not adequately connected inside the monitor, and data from these monitors were excluded from further analysis. During the subsequent identification and removal of outliers (values < Q1 − 1.5 × IQR or > Q3 + 1.5 × IQR) from the colocation data, <1% of the original data points were removed. Although the monitors measure both O₃ and NO₂, monitor-derived NO₂ correlated poorly with NO₂ measured by the reference analyzer (R²: 0–0.6 across sensors, with all but one value < 0.33) and NO₂ was not further examined or used as part of the calibrations.

After removal of outliers and erroneous data and after integrating by hour, unique calibration constants (slope, intercept) were derived for each monitor (Table 1) using linear regression of the monitor concentrations on data from the reference analyzer during periods when all instruments were collocated at PKU. During calibrations such as these it is important to collect data over a wide range of concentrations and especially over the range of concentrations and environmental conditions one might expect to see over the course of the study. During the ∼3 days of colocation, O₃ measured by the reference analyzer ranged from <1 ppb to 142 ppb. This is comparable to the values measured during home monitoring, which ranged from 2–165 ppb; four data points fell outside of the range of concentrations measured during colocation. Conditions were slightly warmer and with higher humidity (t-test, p < 0.05), on average, during the pre-colocation (mean: 28 °C, 75%) than during the post-colocation (mean: 26 °C, 66%) although it was generally warm and humid during both periods (pre, range: 22–36 °C, 50–100%; post, range: 21–33 °C, 48–90%). There was substantial overlap between environmental conditions measured during the colocations and those encountered indoors and outdoors at homes during home monitoring (hour averages across homes, indoor: 28–38 °C and 36–72% RH; outdoor: 25–46 °C and 21–86% RH), and we would not expect the slightly higher temperatures and lower humidity encountered infrequently during home sampling to have a large impact on our results.

Table 1 Parameters for linear calibrations (monitor = slope × reference + intercept) between monitors and a reference analyzer while all instruments were collocated on a rooftop at Peking University for two days prior to, and one day following, sampling at homes. (RMSE = root mean square error; NRMSE = normalized root mean square error; MAE = mean absolute error)

Monitor use	Pre-colocation			Post-colocation			Combined colocation data (pre + post)					MAE (ppb)
	Intercept (ppb)	Slope	R ^2	Intercept (ppb)	Slope	R ^2	Intercept (ppb)	Slope	R ^2	RMSE (ppb)	NRMSE (%)
a Sensor was not working properly. Note: data from home 6, indoor, were excluded due to sensor malfunction.
Base station #1	−21	0.6	0.66	−16	1.0	0.87	−14	0.6	0.57	30.8	61	21.7
Base station #2	21	1.4	0.83	24	1.9	0.94	28	1.4	0.83	16.8	33	14.0
Home 1 – indoor	0	1.3	0.65	18	1.5	0.76	14	1.2	0.64	26.9	53	20.6
Home 1 – outdoor	35	1.0	0.81	33	1.0	0.96	34	1.0	0.86	15.1	30	11.8
Home 2 – indoor	−12	1.4	0.87	0	1.4	0.98	−4	1.3	0.88	13.8	27	11.2
Home 2 – outdoor	−16	1.4	0.84	−7	1.3	0.97	−11	1.3	0.87	14.1	28	11.2
Home 3 – indoor	17	1.3	0.83	26	1.2	0.91	22	1.2	0.85	15.3	30	12.1
Home 3 – outdoor	−13	1.3	0.88				−13	1.3	0.88	14.3	28	12.4
Home 4 – indoor				11	1.2	0.97	11	1.2	0.97	3.6	7	2.2
Home 4 – outdoor				−3	1.6	0.97	−3	1.6	0.97	4.8	9	3.7
Home 5 – indoor	−9	1.5	0.90	13	1.4	0.98	4	1.4	0.89	13.1	26	10.7
Home 5 – outdoor	9	0.8	0.80	3	1.1	0.92	8	0.8	0.82	17.2	34	13.5
Home 6 – outdoor	2	1.4	0.89	12	1.4	0.97	7	1.4	0.91	11.6	23	9.1
Personal #1	0	0.9	0.85	23	0.9	0.37	14	0.8	0.69	24.9	49	21.4
Personal #2	−23	0.6	0.83	1	0.2	0.95	−12	0.4	0.76	21.2	42	18.0
Personal #3	−7	0.6	0.82	9	0.6	0.83	3	0.6	0.72	23.5	46	19.3
Average	−1	1.1	0.82	10	1.2	0.89	6	1.1	0.82	16.7	33	13.3
Standard deviation	17	0.4	0.08	14	0.4	0.16	15	0.4	0.11	7.4	14	5.8

---

During pre-colocation, the intercepts ranged from −23 to 35 ppb (mean ± SD: −1 ± 17 ppb), and the slopes ranged from 0.6 to 1.5 (mean ± SD: 1.1 ± 0.4) (Table 1). During post-colocation, the intercepts ranged from −16 to 33 ppb (mean ± SD: 10 ± 14 ppb), and the slopes ranged from 0.2 to 1.9 (mean ± SD: 1.2 ± 0.4). Coefficients of determination (R²) for monitor vs. reference analyzer data were 0.56–0.90 during pre-colocation and 0.36–0.98 during post-colocation. A paired t-test was used to compare, by monitor, the intercepts and the slopes between the pre- and post-colocations; differences in the slopes were not statistically significant (p > 0.05) although differences in the intercepts were. Some monitors had colocation data from only one of the two periods due to issues of poor sensor connection or for logistical reasons. As such, final calibration constants were derived after aggregating pre- and post-colocation data for each monitor and conducting linear regression between these data and data from the reference analyzer. Although the slopes of the individual monitors did not change appreciably over the several weeks of the study, over longer periods of time in these highly polluted environments this may not hold true, especially if environmental conditions change during this time due to a change in season. Furthermore, previous work by others has indicated that over multiple months of use these Alphasense electrochemical sensors may experience drift, although the nature of the drift may not be uniform across sensors thus indicating the need for careful sensor-by-sensor recalibration over time.²⁰

The parameters derived from regression were used to calibrate all data collected by the monitors during colocation, during home sampling, and for personal exposure. Monitor errors were calculated as the difference between the calibrated sensor values and data from the reference analyzer during the colocation periods (Table 1). After applying the calibration, root-mean-square deviations (RMSE) were calculated to measure the differences between the predicted values and observed values. The average RMSE was 16.7 ppb, and the mean absolute error (MAE) was 13.3 ppb. This RMSE is comparable to what was found in China in June by others using the same Alphasense O₃ sensors (14.7 ppb).¹⁹ Others have previously found slightly lower RMSE and MAE values for Alphasense O₃ sensors using larger data sets and including more parameters for calibration, although some of these differences may relate to the conditions experienced in these environments which may not reflect the conditions typical in urban China. For example, in Cross' examination of these sensors in Boston, USA, a prolonged colocation (4.5 months) led to the collection of more than 25 000 data points (25% used for model training), for an RMSE of 9.71 ppb and a MAE of 7.34 ppb, although the examination was of an older model of the O₃ sensor (OX-B421).²¹ Employing a machine learning approach, Zimmerman et al. obtained a MAE of 3.36 ± 0.41 ppb in Pittsburgh, USA.²² On average, our R² values were consistent with those measured previously (mean: 0.82),^7,22 although some individual sensors in our study had low R² values (i.e., 5 sensors <0.8). Taking into account the average concentration during the calibration periods (51 ppb) the normalized root mean squared error (NRMSE) in our study was 33%, with sensor-specific values ranging from 7 to 61%. If we exclude sensors that had an R² < 0.7 during either of the two calibration periods (n = 3), our metrics improve to a mean R² of 0.86, RMSE of 14.2 ppb, NRMSE of 28% and MAE of 11.5 ppb. Substantial variability in the values for these metrics across sensors highlights the importance of conducting sensor-specific evaluations prior to use in the field as performance is quite variable even when sensors are brand new; such evaluations may allow researchers to identify and remove from further use any sensors that are under-performing.²¹

As we left two monitors (base station #1 and #2 in Table 1) alongside the reference analyzer for the duration of the study (n = 21 days) we were able to evaluate calibrations derived from a prolonged colocation against the 3 day colocation used for calibration in this study. This longer colocation captured a broader range of concentrations of O₃ (up to ∼250 ppb) than was seen during our 3 day colocation. The RMSE for the two monitors based on the prolonged colocation/calibration were 37.3 and 17.9 ppb – slightly higher than the 30.8 and 16.8 ppb obtained for these same monitors when evaluation was restricted to the 3 days of colocation. Similarly, R² for the prolonged colocation was 0.58 (3 day: 0.57) and 0.80 (3 day: 0.83). This is comparable to work done by others with these same Alphasense O₃ sensors in which they found an R² of 0.65 during a 5 day colocation period used for sensor calibration.¹⁹ In sum, in this particular instance conducting a substantially longer colocation does not appear to improve data quality. Our 3 days of colocation already capture a broad range of concentrations and environmental conditions and although the longer colocation included higher concentrations of O₃, few measurements at homes (n = 4) exceeded the concentrations measured during the 3 days of colocation anyways.

As is evident in Fig. 4, the monitors seem to generally track the O₃ data measured by the reference analyzer and to pick up the diurnal trends in concentrations; the monitors correlate well with each other (R²: 0.82). However, there are numerous periods during which the monitors substantially overestimate or underestimate the reference analyzer; one monitor experienced spikes and drops in concentrations (see Fig. 4, ∼Aug 3) that were either attenuated in the data measured by the other monitor and the reference, or which did not match up at all with these data. In previous work with Alphasense oxidant sensors, concentrations of O₃ were underestimated by up to 187% when there was a mix of NO₂ and O₃,¹⁶ and overestimated by 20–40% when concentrations were at their peak.²³ The differences we observed between the reference and our sensors may thus reflect, in part, diurnal changes in the concentrations of NO_x,. The marked differences in the concentrations measured by the reference analyzer and the sensors point to the need for a greater investigation of calibration methods and environmental conditions or cross-sensitivities to other pollutants that may be biasing these monitor-based measurements in this environment. This also highlights the need for careful comparison of measurements collected by different sensors following collocation as differences between measurements at different locations may fall within the error of the two sensors used to make those measurements.

Fig. 3 (Left) Hour-averaged data from the initial (top) and final (bottom) colocation periods for all monitors (coloured lines = calibrated monitor data, black line = reference analyzer data); (Right) hour-averaged, calibrated indoor, outdoor and personal exposure data from all homes.

Fig. 4 Hour-averaged data from the reference analyzer (black) and from the two monitors collocated with the reference analyzer at PKU for the entire field campaign.

3.1.2 Limit of detection. Monitor-specific LODs ranged from 3.3 to 12.6 ppb, with a mean of 7.0 ppb (median: 6.2 ppb). Others have previously identified the LOD for Alphasense O₃ sensors as 1.8 ppb (OX-B421)²⁴ and other low-cost O₃ sensors were found to have LODs ranging from 0.5–23.4 ppb;²⁵ our estimates align with the LODs previously determined for similar technology. Additional work should be done in future studies to confirm that this LOD is consistent across monitoring conditions. For monitors that lacked data during the periods of low reference O₃ concentrations (n = 2) and for which it was not possible to determine a monitor-specific LOD, the mean LOD from all sensors was used in place of a sensor-specific LOD.

3.1.3 Effects of temperature and relative humidity on ozone concentrations. According to the sensor manufacturer, the electrochemical gas sensors have high selectivity, sensitivity, and dynamic range, and contain a sulfuric acid electrolyte for ionic transport.²⁶ The electrolyte loses water in low humidity and gains water in high humidity, and the sensors are specified to operate in the range of 15% to 85% RH.²⁷ Home-specific mean hourly relative humidity during monitoring at homes ranged from 53 to 57% indoors and 50 to 59% outdoors, with hourly values from 36 to 72% indoors and 21 to 87% outdoors. We assessed whether inclusion of temperature and humidity data in the calibration would improve sensor performance. Home-specific mean hourly temperatures during monitoring ranged from 30 to 33 °C indoors and 32 to 35 °C outdoors, with hourly values ranging from 28 to 38 °C indoors and 25 to 46 °C outdoors. Calibrated O₃ data from all monitors were positively correlated with temperature, as recorded by the monitors, with an average correlation coefficient of 0.78. As expected, RH recorded by the monitors was inversely related to the calibrated O₃ concentration (r = −0.67). After calibration of monitor data using the O₃ from the reference analyzer, model residuals were poorly correlated with temperature (R² = 0.08) and RH (R² = 0.11), and showed no strong pattern when plotted against these predictors indicating that inclusion of these additional variables in the calibration model would likely have little effect on model performance. However, in a sensitivity analysis we did include temperature and RH in the model; inclusion of these predictors led to a slight reduction in the mean RMSE across monitors – from 16.7 ppb in the basic model (calibration against reference O₃ only) to 14 ppb – but this required the exclusion of data from sensors in which the temperature/RH sensor was not working. Given the marginal gains in performance at the cost of excluding data from some sensors to achieve these gains, temperature and humidity were ultimately excluded from the calibration in this study.

3.2 Outdoor, indoor, and personal ozone concentrations

3.2.1 Ozone outdoors. Although mostly sunny during the field campaign, there was some rain throughout this study. This is typical of Beijing this time of year.²⁸ There was substantial variation in ambient concentrations of O₃ between homes during sampling (Fig. 3 and 5), with home 1 having the highest mean concentration (46 ppb), home 2 having the lowest (12 ppb), and some homes (i.e., homes 1, 3 and 5) routinely exceeding hourly averages of 50 ppb while others seldom (home 4) or never (home 2) did. As expected, diurnal patterns in ambient O₃ were visible for all homes, with concentrations dropping at night to values <10 ppb for all homes at some point during monitoring. Hour-averages from urban monitors operated by the China Environmental Monitoring Center over the duration of in-home sampling were similarly variable, having monitor-specific means ranging from 60 to 100 ppb, standard deviations from 59 to 71 ppb, minima of 4 ppb or less, and maxima from 269 to 335 ppb.²⁹

Fig. 5 Distribution of hour-averaged indoor, outdoor and personal concentrations of O₃ by home; box limits indicate first and third quartiles (median = horizontal bar), whiskers extend no more than 1.5 times the interquartile range from the box and points beyond this are indicated as dots; means are indicated by an “x”.

Significant spatial variation can exist in O₃ concentrations across small (i.e., indoor vs. outdoor) and larger (i.e., between sampling sites) distances.³⁰ Where the homes were located likely impacts the concentrations of O₃ – for example, home 1 was in an area that was in close proximity to traffic, near several large commercial centers and companies at the third-ring road. Home 5 was also next to a busy road, and there was a construction site behind the apartment where the monitor was located. Previous work in Beijing found lower O₃ concentrations at traffic sites versus other urban sites, although the exact definition of a “traffic site” (e.g. proximity to the roadway) was not clear from the paper.³¹ Although home 1 is in a neighborhood close to a major roadway, the home itself is set back in the neighborhood. In spite of a relative proximity to main roads, it is possible that homes 1 and 5 are also close to sources of volatile organic compounds (VOCs) leading to their higher concentrations of O₃. For example, at home 1, nearby restaurants and an automobile service centre may be emitting VOCs and for home 5, emissions may be coming from a nearby cleaner. As we did not methodically identify these types of nearby pollutant sources during the field campaign, we are unable to determine what these sources are with certainty. Home 4 may have experienced lower ozone concentrations due to nearby sources of NO_x and fewer sources of VOCs. Although both on the Tsinghua campus, there were large and significant (p < 0.05, paired Wilcoxon sign rank test) differences in the concentrations of O₃ measured at home 2 (median: 11.0 ppb, range: 3.0–32.2 ppb; sensor RMSE: 14.1 ppb, MAE: 11.2 ppb) and home 3 (median: 37.8 ppb, range: 9.0–121.1 ppb; sensor RMSE: 14.3 ppb, MAE: 12.4 ppb). These homes were located on different parts of the campus which are likely influenced by differences in NO_x emissions that influence O₃ concentrations. The mean difference in concentrations between these two homes was 27.3 ppb (median 19.5 ppb), with differences ranging from −7.0 to 118.1 ppb or 44.9–190.3% (relative percent difference); we believe these to be real differences in concentrations between the two homes although the exact magnitude of the difference may be slightly more or less than indicated due to the error associated with these measurements.

There are numerous additional ambient monitors in Beijing, and we compared values from our low-cost ambient monitors to data from the nearest monitors. Our monitors showed similar trends but substantially lower values than those from the nearest reference monitors. As the distance between the closest regulatory monitor and home monitors ranged from ∼0.6–3.5 km we might expect to see differences between the concentrations in these environments. However, the systematically lower values measured by the monitors as compared to the reference data may be due to a short-coming of the technology itself; previous work with Alphasense sensors for measuring O₃ has shown that when sensors are exposed to a mixture of O₃ and NO₂ – as is common when collecting ambient measurements due to traffic-related pollution – O₃ concentrations are often underestimated.¹⁶ In Beijing, additional sources of NO_x in the summer may include fossil fuel combustion for manufacturing, construction, and for the generation of electricity.³²

3.2.1 Ozone indoors. Although outdoor concentrations were, in general, substantially higher than indoor concentrations (Fig. 3 and 5), this was not true for home 2, which was identified as being non-compliant with respect to keeping doors and windows closed during sampling. With the exception of home 2, these indoor–outdoor differences were significant overall (p < 0.05, paired Wilcoxon signed rank test). Similarly, the standard deviations for concentrations of O₃ outdoors were larger than for O₃ indoors in all but home 2, highlighting variation in concentrations outdoors that is not seen indoors. Forty-four percent of all data points from home (indoor/outdoor) and personal monitoring fell below the monitor-specific LODs, comprising 77% of all indoor monitoring data and 62% of personal exposure data, although these generally low concentrations of O₃ indoors are not unexpected as O₃ rapidly reacts with surfaces.³³ Given the variation in sensor performance, it is important to consider the magnitude of these differences in the context of the sensor-specific performance and errors; for indoor–outdoor differences, the differences are large enough (concentrations indoors are often < LOQ and outdoors are up to 150 ppb) during most of the day that we believe these to be real differences in concentrations. It is possible that some of the lower ambient concentrations measured (e.g. those measured at night, for the most part) do not differ significantly from the indoor measurements due to sensor error in the measurements.

No homes had central heat or ventilation units, although some homes did have an air conditioner and all participants indicated the air conditioner was used some or most of the time. Additional information about the homes has been previously published.¹⁴ In addition, high summer temperatures led to some participants opening windows to reduce temperatures indoors. Compliance was initially determined by examining the simultaneous indoor and outdoor pollutant concentrations at the home. It was readily apparent in the data when indoor concentrations closely mirrored ambient concentrations in terms of trends in concentrations but, more importantly, also in the magnitude of the concentrations. While indoor concentrations of O₃ tend to be quite low with the doors and windows closed, the concentrations are higher when the doors and windows are open providing a relatively constant source of O₃. These indicators of non-compliance in the data were further supported by visits to home 2 during which we noted that the patio door and/or windows between outdoors and the bedroom were open when we arrived. While homes 2 and 5 experienced many hour-averaged values above 10 ppb, most concentrations of O₃ indoors at other homes were quite low (<10 ppb), and for a number of homes (homes 1, 3 and 4) indoor concentrations fell almost exclusively below the LODs. The higher concentrations indoors at home 5 may indicate that the doors and/or windows were opened periodically in this home, although not as often as home 2, leading to greater concentrations of O₃ indoors than were seen in most other homes; it is also possible that this home was less well-sealed than the other homes sampled.

3.2.2 Personal exposure to ozone. The large variability in ambient ozone concentrations measured simultaneously across a highly urban region and specifically at homes, and the changes in both indoor and outdoor values over time illustrate the importance of considering activity patterns when determining exposures to O₃. Measurements from a few stationary ambient monitors may poorly reflect personal exposure if an individual is moving about a city over the course of the day and moving between indoor and outdoor environments.³⁴ As participants carried personal exposure monitors with them throughout their normal daily activities, we encountered more erroneous data during these sampling periods than we did with the indoor/outdoor monitoring likely due to the disconnection of electronic components during participant movement. In the event of sensors becoming disconnected from the other components of the monitoring package (e.g., the printed circuit board), the monitoring package continued to log values, but these values have no logical interpretation. It is for this reason that these data have been classified as “erroneous” rather than missing – gaps in the data only existed if the entire monitoring package was shut off or not functional (i.e., due to a missing SD card for recording data, disconnection from a power source). This resulted in usable personal exposure data for only three participants.^1,2,6 While this reduced our already limited sample size, it provided us with an opportunity to understand the limitations of these monitors and to refine them prior to the similar, but larger, field campaign in Shanghai. To reduce data loss, a method to more firmly connect the sensor within the package is warranted.

Others have previously measured the response times of Alphasense O₃ sensors to various concentrations of O₃ and found them to be two minutes or less.³⁵ This lagged response may lead to some error as participants move between environments with drastically different concentrations of ozone (i.e., moving between indoor and outdoor environments), although for this study data were ultimately hour-averaged and this small lag is thus less of a concern. In addition, if environments differ drastically in terms of humidity and sensors are exposed to a change in RH of 20% min⁻¹ or more they can experience a shock response and may exhibit an artificial spike in output signal; it typically take up to 40 minutes to recover from such a shock.²³ In contrast, gradual changes in RH appear not to appreciably impact sensor output.²³ During personal exposure monitoring, <0.03% of minute readings were associated with a change in RH > 20% min⁻¹ and the mean change in RH was 0.004% min⁻¹; in the current dataset, we think it unlikely that the shock response mentioned is biasing our results.

Personal exposure data from homes 1 and 6 were lower, on average, than outdoor concentrations (Fig. 5), and these differences were significant (p < 0.05, paired Wilcoxon signed rank test). Home 2 had the highest personal exposure even though this home had the lowest outdoor concentrations, and differences between personal exposure and concentrations measured indoors and outdoors were not significant. The differences noted for homes 1 and 6 may relate to the specific activities that the individual was doing during personal monitoring – the outdoor O₃ monitor reflects the concentrations at the individual's home, but the individual may have been exposed to higher or lower ozone concentrations during their commute, while at work, in other rooms in the home without air purification, or while they were doing other activities outside. In addition, personal care and other products and materials used in the home are known sources of volatile organic compounds, which are ozone precursors. It is possible that the personal monitoring captured some generation of ozone due to these products. During personal monitoring we asked participants to complete a brief time-activity form indicating whether they were at home, at work/school, or outdoors each hour. The participant in home 1 spent, on average, 13.25 hours in their bedroom, 4.25 hours in other rooms at home, 6 hours at the office, and 0.5 hours outside per day. The participant at home 5 spent 19 hours in the bedroom, 4 hours in other rooms at home and 1 hour outside, and the participant at home 6 spent 14 hours in the bedroom, no time in other rooms at home, 8 hours at the office, and 2 hours outside. It is likely that even within these environments there were gradients in O₃, and the collection of more detailed time-activity data would facilitate the interpretation of the contributions of different microenvironments to overall exposure. Furthermore, the personal exposure data readily highlight times when the participant was outdoors and away from the home, where concentrations of O₃ were higher than those measured indoors or directly outside the home (see Fig. 3 – home 6). Data like these – that are not consistently reflected in data from stationary monitors – help us to better understand the minutiae of personal exposure and which environments may be contributing more to exposure.

4. Strengths, limitations and further research

As a small pilot study, this work has numerous limitations. First and foremost, the sample size (6 homes) and duration of the study (∼2 weeks in homes) allowed us to gain a preliminary understanding of the concentrations of O₃ in Beijing but limit any broad conclusions about O₃ in these types of highly polluted urban environments in China. Second, due to variation in the performance of the sensors even when they are brand new, differences in calibrated measurements cannot be taken at face value. Measurements must be carefully examined in the context of the performance of the sensor(s) being used to make the measurements.

A strength of this work is that it highlights the potential of these monitors, once calibrated and when evaluating measurements in the context of sensor performance, for capturing spatial and temporal variability in O₃ concentrations; sensors captured substantial differences in ambient concentrations between homes located in different parts of the city, daily variability in ambient O₃ concentrations, and stark contrasts between indoor and outdoor concentrations. This study also illustrates a potential alternative use for these sensors – to monitor compliance in a study such as this in which it is important to keep doors and windows closed (for example, when evaluating the maximum benefit of a filtration intervention). In our data, high concentrations of O₃ indoors were a good indicator of a poorly sealed home or one in which windows and doors were routinely opened.

As these are low-cost devices and therefore less accurate out-of-the-box than reference analyzers, it is critical to determine an appropriate method of calibration to maximize data quality. Although the inclusion of temperature and relative humidity did not greatly improve data accuracy in this context, it may do so under different conditions (e.g., in Shanghai in seasons other than summer and for a longer duration); each research study should consider similar evaluations of parameters for inclusion during calibrations. Others have recently examined the use of random forest models to calibrate data from Alphasense O₃ sensors and have seen large improvements in accuracy with these models over linear regression models.²² However, in the aforementioned work the random forest model approach was applied to a much larger (∼6 months of data, using 4 weeks of data for model training) and more complex (containing numerous pollutants, temperature and RH data) dataset; although such an approach could lead to improvements in data accuracy with our monitors, these techniques were not applied to the much smaller dataset used in the current study.

The variation in calibration parameters observed in this study illustrates the need to generate calibration constants for each individual sensor. For this study, a minimum of 2 days of colocation with the reference instrument was required – one day prior to deployment in homes and one day following – to examine potential drift of the sensors. Furthermore, due to the daily pattern in O₃ concentrations (i.e., with concentrations highest during daylight hours and lowest at night), conducting colocations over a period of 24 hours minimum ensured we captured a broad range of concentrations; this would not have been achieved had we collocated for a much shorter period. This approach also allowed us to capture times with the low O₃ concentrations needed to generate appropriate calibration constants for in-home measurements as concentrations of O₃ indoors can be, and were in this study, quite low. Environmental conditions and pollutant concentrations were generally stable during the field campaign and colocations beyond the 3 day colocations used for the majority of monitors in this study led to little improvement in data accuracy. Nonetheless, a prolonged colocation may lead to greater gains under different conditions, if, for example, a few days does not capture the full suite of environmental conditions possible during the study; prolonged colocations should not be ruled out in future research. In short studies such as this the potential benefits of long colocations must also be weighed with restrictions on study resources and time limitations.

Lastly, although we integrated data over an hour, personal exposure monitoring may benefit from integrating data over a shorter period of time, especially if individuals are moving frequently between environments with different conditions. However, this must be balanced by losses in data quality that may occur with increased granularity. In recent work by others using these sensors the R² between sensor data and reference data while both were collocated but in motion was determined to be 0.56 for minute-averaged data; this highlights some of the challenges to data quality that arise when sensors are in motion and data are assessed over fine time-scales.¹⁹ In the current study, it is possible that different averaging periods would have led to different LODs for the sensors, although this was not investigated. This pilot highlighted some of the shortcomings of the design of the monitors, thereby allowing us to modify them to be more robust for the subsequent study in Shanghai; while the components used in these monitors seem to work well, there is a need to improve upon how sensors and other monitoring package components are affixed inside the package housing and to the in-house, custom-designed printed circuit board in order to prevent data loss.

5. Conclusions

Our overall objectives in this study were to evaluate low-cost, time-resolved O₃ monitors in an urban area with relatively poor air quality (i.e. Beijing) for monitoring indoor and outdoor concentrations of O₃ and for use in personal exposure monitoring. Most monitors performed well with high R² between monitors and a reference analyzer with values ranging from 0.66–0.90 during the pre-colocation period (mean = 0.82) and 0.37–0.98 during the post-colocation period (mean = 0.89). After applying calibrations to the monitors, the mean RMSE was 16.7 ppb, the MAE was 13.3 ppb and the NRMSE was 33%, although performance varied substantially between sensors. Monitor-specific LODs ranged from 3.3 to 12.6 ppb, highlighting the need to determine LOD values and quantify performance on a sensor-by-sensor basis. A great deal of variability was seen for outdoor concentrations at the homes across Beijing while the bulk of indoor data (77%) fell below the detection limit of the monitors. Personal exposure reflected, at times, the low concentrations indoors, and at other times contributions that far exceeded the concentrations measured in and near the home. Overall, the monitors show promise for providing a low-cost option for accurately measuring concentrations of O₃ outdoors, where pollutant concentrations tend to be relatively high (i.e., greater than the monitor LODs). However, measurements must be evaluated in the context of the sensor-specific parameters of performance as there is large variation in these parameters. Regardless of differences in performance it is clear that these sensors can be used to identify temporal changes in concentrations of O₃. This study also illustrates some, although perhaps limited, utility for these sensors in environments with lower concentrations of O₃ (e.g., indoors); utility will depend on just how low the concentrations are in these environments. Lastly, in an urban area with relatively high levels of pollution, our results indicate that these sensors may be useful for capturing large differences in O₃ concentrations experienced by individuals as they move between environments. Further iterations of the monitoring package require more robust attachment of the sensor itself within the monitor to prevent erroneous data or data loss.

Conflicts of interest

There are no conflicts of interest to declare.

Acknowledgements

The project team thanks Underwriters Laboratory (UL) Inc. for support and funding. This study was also supported in part by grants from the National Natural Science Foundation of China (51420105010). Thanks to Amway (China) Co., Limited, for lending the air purifiers for use in this study; however, the company had no involvement in study design, implementation, or data interpretation. Thanks to Yusheng Wu from Peking University for their support including access to and data from their rooftop monitoring lab. Thank you also to Linchen He of Duke University, and Wei Jingya of Tsinghua University for their assistance with logistics and data collection in the field, and to the participants in this study for allowing us access to their homes.

References

1. Chinese National Environmental Monitoring Center, July 2016 City Air Quality Status Report, Chinese National Environmental Monitoring Center, 2016 July.
2. China, Ambient Air Quality Standards: GB 3095-2012, 2012.
3. K. He, H. Huo and Q. Zhang, Urban Air Pollution in China: Current Status, Characteristics, and Progress, Annu. Rev. Energy Environ., 2002, 27(1), 397–431.
4. D. E. Williams, G. S. Henshaw, M. Bart, G. Laing, J. Wagner and S. Naisbitt, et al., Validation of low-cost ozone measurement instruments suitable for use in an air-quality monitoring network, Meas. Sci. Technol., 2013, 24(6), 065803.
5. W. Y. Yi, K. M. Lo, T. Mak, K. S. Leung, Y. Leung and M. L. Meng, A Survey of Wireless Sensor Network Based Air Pollution Monitoring Systems, Sensors, 2015, 15(12), 31392–31427.
6. A. C. Lewis, J. D. Lee, P. M. Edwards, M. D. Shaw, M. J. Evans and S. J. Moller, et al., Evaluating the performance of low cost chemical sensors for air pollution research, Faraday Discuss., 2016, 189, 85–103.
7. L. Spinelle, M. Gerboles, M. G. Villani, M. Aleixandre and F. Bonavitacola, Field calibration of a cluster of low-cost available sensors for air quality monitoring. Part A: Ozone and nitrogen dioxide, Sens. Actuators, B, 2015, 215, 249–257.
8. R. M. Duvall, R. W. Long, M. R. Beaver, K. G. Kronmiller, M. L. Wheeler and J. J. Szykman, Performance Evaluation and Community Application of Low-Cost Sensors for Ozone and Nitrogen Dioxide, Sensors, 2016, 16(10), 1698.
9. K. Sadighi, E. Coffey, A. Polidori, B. Feenstra, Q. Lv and D. K. Henze, et al., Intra-urban spatial variability of surface ozone in Riverside, CA: viability and validation of low-cost sensors, Atmos. Meas. Tech., 2018, 11(3), 1777–1792.
10. E. Hall, S. Kaushik, R. Vanderpool, R. Duvall, M. Beaver and R. Long, et al., Integrating Sensor Monitoring Technology into the Current Air Pollution Regulatory Support Paradigm: Practical Considerations, Am. J. Environ. Eng., 2014, 4(6), 147–154.
11. L. Li, Y. Wang, Q. Zhang, J. Li, X. Yang and J. Jin, Wheat straw burning and its associated impacts on Beijing air quality, Sci. China, Ser. D: Earth Sci., 2008, 51(3), 403–414.
12. J. Lang, Y. Zhang, Y. Zhou, S. Cheng, D. Chen and X. Guo, et al., Trends of PM_2.5 and Chemical Composition in Beijing, 2000–2015, Aerosol Air Qual. Res., 2017, 17(2), 412–425.
13. Z. Wang, Y. Li, T. Chen, D. Zhang, L. Li, B. Liu, et al., Science–Policy Interplay: Improvement of Air Quality from 2008 to 2014 in Beijing and the Scientific Approach to Achieve APEC Blue, 2016.
14. K. K. Barkjohn, M. H. Bergin, C. Norris, J. J. Schauer, Y. Zhang and M. Black, et al., Using Low-cost sensors to Quantify the Effects of Air Filtration on Indoor and Personal Exposure Relevant PM2.5 Concentrations in Beijing, China, Aerosol Air Qual. Res., 2018 DOI:10.4209/aaqr.2018.11.0394.
15. M. I. Mead, O. A. M. Popoola, G. B. Stewart, P. Landshoff, M. Calleja and M. Hayes, et al., The use of electrochemical sensors for monitoring urban air quality in low-cost, high-density networks, Atmos. Environ., 2013, 70, 186–203.
16. C. Zuidema, N. Afshar-Mohajer, M. Tatum, G. Thomas, T. Peters and K. Koehler, Efficacy of Paired Electrochemical Sensors for Measuring Ozone Concentrations, J. Occup. Environ. Hyg., 2018, 1–12.
17. L. Spinelle, M. Aleixandre and M. Gerboles, Protocol of evaluation and calibration of low-cost gas sensors for the monitoring of air pollution, Report No.: Report EUR26112EN, European Commission, 2013.
18. Y. Zhan, K. Johnson, C. Norris, M. M. Shafer, M. H. Bergin and Y. Zhang, et al., The influence of air cleaners on indoor particulate matter components and oxidative potential in residential households in Beijing, Sci. Total Environ., 2018, 626, 507–518.
19. L. Chatzidiakou, A. Krause, O. A. M. Popoola, A. Di Antonio, M. Kellaway and Y. Han, et al., Characterising low-cost sensors in highly portable platforms to quantify personal exposure in diverse environments, Atmos. Meas. Tech., 2019, 2019, 1–19.
20. A. Ripoll, M. Viana, M. Padrosa, X. Querol, A. Minutolo and K. M. Hou, et al., Testing the performance of sensors for ozone pollution monitoring in a citizen science approach, Sci. Total Environ., 2019, 651, 1166–1179.
21. E. S. Cross, L. R. Williams, D. K. Lewis, G. R. Magoon, T. B. Onasch and M. L. Kaminsky, et al., Use of electrochemical sensors for measurement of air pollution: correcting interference response and validating measurements, Atmos. Meas. Tech., 2017, 10(9), 3575–3588.
22. N. Zimmerman, A. A. Presto, S. P. N. Kumar, J. Gu, A. Hauryliuk and E. S. Robinson, et al., A machine learning calibration model using random forests to improve sensor performance for lower-cost air quality monitoring, Atmos. Meas. Tech., 2018, 11(1), 291–313.
23. X. Pang, M. D. Shaw, A. C. Lewis, L. J. Carpenter and T. Batchellier, Electrochemical ozone sensors: A miniaturised alternative for ozone measurements in laboratory experiments and air-quality monitoring, Sens. Actuators, B, 2017, 240, 829–837.
24. N. Castell, F. R. Dauge, P. Schneider, M. Vogt, U. Lerner and B. Fishbain, et al., Can commercial low-cost sensor platforms contribute to air quality monitoring and exposure estimates?, Environ. Int., 2017, 99, 293–302.
25. A. C. Rai, P. Kumar, F. Pilla, A. N. Skouloudis, S. Di Sabatino and C. Ratti, et al., End-user perspective of low-cost sensors for outdoor air pollution monitoring, Sci. Total Environ., 2017, 607–608, 691–705.
26. Alphasense Ltd, Alphasense Application Note AAN 106 – Humidity Extremes: Drying Out and Water Absorption, 2013, ch. 12.
27. Alphasense Ltd, OX-A431 Oxidising Gas Sensor: Ozone + Nitrogen Dioxide 4-Electrode, 2016.
28. China Meteorological Administration, The Public Weather Service Center of CMA, 2019, http://www.cma.gov.cn/.
29. China Environmental Monitoring Center, 2019, http://113.108.142.147:20035/emcpublish/.
30. T.-Y. Lin, L.-H. Young and C.-S. Wang, Spatial variations of ground level ozone concentrations in areas of different scales, Atmos. Environ., 2001, 5799–5807.
31. H. Zhao, Y. Zheng, T. Li, L. Wei and Q. Guan, Temporal and Spatial Variation in, and Population Exposure to, Summertime Ground-Level Ozone in Beijing, Int. J. Environ. Res. Public Health, 2018, 15(4), 628.
32. M. Wang, T. Zhu, J. Zheng, R. Y. Zhang, S. Q. Zhang and X. X. Xie, et al., Use of a mobile laboratory to evaluate changes in on-road air pollutants during the Beijing 2008 Summer Olympics, Atmos. Chem. Phys., 2009, 9(21), 8247–8263.
33. C. J. Weschler, Ozone in Indoor Environments: Concentration and Chemistry, Indoor Air, 2000, 10(4), 269–288.
34. S. A. Abdul-Wahab, C. S. Bakheit and S. M. Al-Alawi, Principal component and multiple regression analysis in modelling of ground-level ozone and factors affecting its concentrations, Environ. Model. Softw., 2005, 20(10), 1263–1271.
35. L. Spinelle, M. Gerboles and M. Aleixandre, Performance Evaluation of Amperometric Sensors for the Monitoring of O₃ and NO₂ in Ambient Air at ppb Level, Procedia Eng., 2015, 120, 480–483.

---