Optimization Approaches to Ameliorate Humidity and Vibration Related Issues Using the MicroAeth Black Carbon Monitor for Personal Exposure Measurement

Abstract

Exposure to ambient black carbon (BC) is associated with adverse health effects. BC levels display large spatial and temporal variability in many settings, such as cities and rural households where fossil fuel and biomass, respectively, are commonly burned for transportation, heat, and cooking. This article addresses the optimization of the miniaturized personal BC monitor, the microAeth® for use in epidemiology studies. To address false positive and negative peaks in real time BC concentrations resulting from changes in temperature and humidity, an inlet with a diffusion drier was developed. In addition, we developed data cleaning algorithms to address occasional false positive and negative fluctuations in BC readings related to physical vibration, due in part to both dirt accumulations in the optical inserts and degraded components. These methods were successfully used to process real-time BC data generated from a cohort of 9–10 year old children (N = 54) in New York City, who wore 1 or 2 microAeth units for six 24 h time periods. Two-hour and daily BC averages after data cleaning were consistent with averaged raw data (slopes near 1 with R = 0.99, p < .001; R = 0.95, p < .001, respectively), strongly suggesting that the false positive and negative excursions balance each other out when averaged for at least 2 h. Data cleaning of identified suspect events allows more confidence in the interpretation of the real-time personal monitoring data generated in environmental exposure studies, with mean percent difference <10% for 19 duplicate deployments.

Copyright 2013 American Association for Aerosol Research

1. INTRODUCTION

A large proportion of the world's population is exposed to black carbon (BC) either derived from fossil fuel combustion in urban centers or by biomass combustion in rural settings. BC particles are of increasing interest for their climatic, environmental, and widespread health effects (Andreae 2001; Jacobson 2001; USEPA 2012). Though BC typically comprises only a fraction of the fine particulate mass (PM_2.5), recent studies have suggested that certain adverse health effects can be more strongly associated with BC than PM_2.5 (Bell et al. 2009; Patel et al. 2009; Spira-Cohen et al. 2011). The association of BC with negative health effects may simply be due to the fact that BC is a good indicator/tracer of incomplete combustion sources, that is, some other unreported pollutants derived from incomplete combustion can be the mechanistic cause for the associations.

There are common misconceptions within some research communities between elemental carbon (EC) and BC measurements and as such it is worth giving a brief comparison. It is generally thought that airborne BC and EC refer to the same materials: highly condensed refractory carbonaceous particles derived from incomplete combustion. However, both BC and EC are empirically defined by the different methods used for measuring them. If thermal- and/or chemical-based methods are used, then the term EC is applied; if an optical method is used, leveraging the strong light absorption features of these particles, then the term BC is applied. Levels of BC and EC are typically highly correlated, but the slope of the best-fit line can vary by location, season and laboratory, with analytical problems being able to occur with either method (Jeong et al. 2004; Venkatachari et al. 2006; Quincey et al. 2009). In general, optical methods have been more widely used, due to their ease of use, relatively low analytical cost, nondestructiveness, and the association with the optical properties of the particles in the atmosphere. Real-time optical sensing of BC underlies the operating principle of the microAeth® (Aethlabs, San Francisco, CA, USA) instrument which is the subject of this article.

Personal monitoring is widely viewed as the gold standard for exposure assessment of air pollutants, especially of those with large spatial and temporal variations or special exposure pathways. In spite of its recognized advantages, the use of personal monitors as an exposure tool in large-scale epidemiological studies has been extremely limited to-date. This reflects the size, weight, noise, and cost associated with using available personal monitoring technology. Due to these constraints, most personal monitoring studies thus far have been small (Spengler et al. 1985; Chillrud et al. 2004). The microAeth is lightweight, compact in size, and easy to operate; thus, it has the potential to be used in epidemiological studies as both a personal and indoor air monitor. In addition, as a real-time personal monitor, the microAeth is providing additional information about exposure patterns and peak levels of exposure.

Even though several hundred microAeth units are in use around the world (Dons et al. 2012), only limited validation data have been presented in the literature (Hagler et al. 2011; Cai et al. in review). In addition, faulty BC readings related to rapid changes in RH and T have been observed with recent field tests (Cai et al. in review), similar to those generally seen for other optical instruments such as particle counters and nephelometers. Furthermore, we and others have observed positive and negative excursions in the BC time series data related to physical movement and vibration, especially after sustained use of the monitors. Here, we report on the development of optimization methods to help overcome RH and T effects and vibration impacts on the BC data by conducting a series of experiments as well as analyzing duplicate data from a field cohort study; recommendations are made about hardware maintenance that can decrease the impact of physical vibrations on the BC data, and quality control algorithms are presented to allow confident use of the real-time data.

2. THEORY OF MICROAETH IN MEASURING BC

The microAeth uses similar principles and algorithms as the full size Aethalometer® (Magee Scientific, Berkeley, CA, USA). They both measure the transmission of light at 880 nm through the active area of filter (referred to as the sensing channel) on which aerosol is collected, and both instruments normalize the sensing channel results to changes in a reference channel, which monitors an unloaded portion of the same filter (Ruoss et al. 1992). A major difference between the Aethalometer and the microAeth is the filter arrangement; the Aethalometer uses a spool of filter material, which can extend measurements for many weeks to months, whereas the microAeth collects particles onto a single filter, which is changed manually before the filter becomes optically saturated.

The primary assumption of both the microAeth and the Aethalometer is that the increase of optical attenuation is proportional to the increase of mass loading of BC on the active area of the filter (Hansen et al. 1984; Hansen and Novakov 1990). The optical attenuation is the natural log of the ratio of the sensing beam detector output (Sen) to the reference beam detector output (Ref). See the online supplementary information (SI) (Section I) for a detailed description of the microAeth and equations describing the calculations.

Theoretically, the sensing signal would decrease continuously with deposition of BC while the reference channel (which has no active flow) would remain constant. However, any changes to the system (e.g., variations in light source, detector, and/or humidity of filter tape) can result in altered readings over time. The method assumes that any system changes will affect both reference and sensing channels in the same way and thus be cancelled out in the calculation.

3. METHODS

The microAeth® units tested in this study were AE51 models. Each weighs about 250 g with a size of 11.7 cm length × 6.6 cm width × 3.8 cm depth. Initial testing of the units was carried out with firmware/software (version S0) that kept the LEDs and detectors on continuously. To extend the runtime on the internal rechargeable battery to more than 24 h, a firmware/software upgrade was made which turned the light source and detectors on and off to save on power (version S1). Firmware S1 was later upgraded to firmware S2 by using prototype code that allowed for the acquisition and averaging of more measurement samples of the optical signals before calculation of BC concentrations to reduce the noise of baseline. The majority of field data were collected using the firmware/software version S1 or S2 and set-up to make BC readings at 5 min intervals and a flow rate of 50 ml/min on a full charge, allowing typical run times of 26–30 h. Some experiments were collected at 1 min frequencies. For each sampling event in New York City (NYC), a new filter made up of Teflon coated borosilicate glass fiber was used.

3.1. Humidity Effects

An environmental chamber experiment was conducted to observe the microAeth under conditions of abrupt environmental changes ([RH] and [T]). This experiment was carried out in the Controlled Environmental Facility (CEF) at Rutgers University, where diesel particles produced by a generator (Model YDG 5500E, Yanmar Inc. Adairsville, GA, USA) were diluted and then introduced into a temperature- and humidity-controlled stainless steel chamber (2.2 m high × 4.1 m wide × 2.7 m deep) (See SI Section I for more details). The principal investigator entered in and out of the chamber with preset levels of RH, T, and soot levels, wearing five microAeth® units that were connected with either regular conductive inlet tubing (Freelin-Wade Co., McMinnville, OR, USA, 1J-425-01) or in-house designs of diffusion drier boxes. The diffusion drier was inline with the regular inlet tubing and comprised of Nafion® (Perma Pure LLC, Toms River, NJ, USA) tubing encompassed by a sealed plastic box containing silica gel bags. Two sizes of the diffusion drier inlet were used in this experiment; the “short” version had 2.7 cm of exposed Nafion tubing and 4 g of silica gel (inside cotton packets), while the “long” version had 5.7 cm of exposed tubing and 6 g silica gel.

Three of the five microAeth units were used to measure BC, while two were used with additional inline sensors to measure the RH and T (HOBO®, ONSET, H08-003-02, accuracy of RH ± 2%, T ± 0.5°C) of the air before and after a diffusion drier. For the three microAeth units that measured BC, one had only regular conductive inlet tubing while the other two had the same conductive inlet tubing but also incorporated in line a diffusion drier that either had a short or a long diffusion drier box.

Initially, the chamber had diluted diesel levels set to about 65 μg/m³, but to avoid overloading the filters at such a high BC level, the source input (diesel generator) was decreased to a point where the chamber concentration was about 12 μg/m³ (based on a nephelometer calibrated to diesel soot) and then RH and T were raised to >90% and ∼38°C, respectively. In contrast, T outside the chamber was 24°C, RH was about 40%, and BC was around 1 μg/m³. After three cycles, that is, walking into, waiting about 15 min and then walking out of the chamber, the temperature level of the chamber was turned down, which caused super saturation of chamber air. Because of super saturation, the real-time display on the nephelometer indicated that levels were increasing dramatically, resulting in the operator turning the diesel source down to essentially zero. After realizing the readings in the nephelometer were false due to condensation, the diesel source was increased again and one more in/out cycle in the chamber was carried out.

3.2. Duplicate Sampling and Vibrational Effects

Duplicate sampling events were conducted when the microAeths were relatively new and again after 17 months of heavy use. When the microAeths were relatively new, two microAeths were worn in a double lined vest on a dry autumn day for about 9 h using only regular inlet tubing during the event. After 17 months of heavy use, a duplicate personal deployment experiment was carried out after first cleaning the optical inserts of the monitors to explore sensitivity to physical vibrations under normal wearing conditions: both units had short drying inlets (to remove chances of RH and T impacts) and the units were worn in a double lined vest. When the units were relatively new the duplicate sampling included time spent in the following microenvironments: roughly 1.5 h in an office (urban area), 0.5 h outdoors (urban area), 1.5 h commuting from the urban area to a suburban area, and about 5.5 h of suburban indoor time including about 15 min cooking. A similar routine was recorded during duplicate personal sampling carried out after 17 months of heavy use, except it also included 1 h of suburban outdoor time during which 45 min of scripted physical activities (jogging, jumping, and running) were made (to investigate the impact of a range of normal and/or extreme activities on measured BC levels) and an 8-h period as fixed site samplers (i.e., stationary).

The Pearson correlation coefficient (R) and the percent difference (%Diff) between the duplicate units were used as indices of quality (i.e., consistency and reproducibility) of the personal data. The %Diff refers to a ratio of difference between two microAeth BC readings (A–B) to the “true” reading (T), in a same time period. Because the true value is not known, the average value on two readings of two duplicate units was selected as the ‘‘true’’ value (T), that is, average (A and B).

Shaking tests and vibrational effects: The response of microAeth units to motions was examined further by a side-by-side testing. Two microAeth using regular inlet tubing were set to record BC data every 1 min at a flow rate of 100 ml/min. Both units were started with a stationary sampling phase for the first 1.75 h, and then a single unit at a time was picked up and shaken for about 10 min, with about 1 h of resting time until the next unit was chosen for this simple shake test. The data for each unit during and after the time that is being shaken can thus be compared directly to the nonshaken unit. Since these shaking tests were conducted in a lab setting with stable RH and T, diffusion drier inlets were not used.

3.3. Development and Validation of Data Cleaning Methods

Personal BC data were collected from 54 children (9–10 years) who lived in Northern Manhattan and the South Bronx. Each child wore a microAeth unit in a vest pocket for six separate 24 h time periods; the inlet tubing ran inside of the double-lined sampling vest and came out at the breathing zone and included a short diffusion drier. Data from 317 deployments on children were available. Duplicate sampling occurred in 21 of these deployments where the subjects wore a vest with two monitors. However, only 19 of the duplicates were successfully completed; the other two duplicate events had a failure. Consequently, data from 317 + 19 or 336 units, in total, were available for further data analysis. Almost all of the field data were collected at a flow rate of 50 ml/min and integration time of 5 min. Other sensors inside the vest included a GPS data logger (Model BT-Q1000eX, Qstarz International Co., Ltd., Taipei, Taiwan) and an actiwatch® (Mini Mitter Inc. Oreqon, OR, USA). All subject-based sampling followed protocols approved by Columbia IRB including informed consent/assent protocols.

Data from duplicate personal sampling were used to validate the methods for data cleaning. Both cleaned and raw data were averaged for a range of time periods (10 min, 20 min, 30 min, 1 h, 2 h, 4 h) for comparisons between the two duplicate units (e.g., raw1 vs. raw2; cleaned1 vs. cleaned2, etc.). The Pearson correlation coefficient (R) and the %Diff between units deployed in duplicate (N = 19) was used to estimate the impact of our methods for data cleaning on data quality (i.e., consistency and reproducibility). In addition, data from each microAeth deployed successfully were compared for raw versus cleaned data (N = 336) for different averaging times and to understand the frequency of events identified by the data cleaning algorithms.

FIG. 1 Time series of BC from the environmental chamber experiment, showing just the three units that measured BC, one with an inlet with regular tubing, one with an inlet that included a “short” diffusion drier, and the third with an inlet that included a “long” diffusion drier (see text).

4. RESULTS AND DISCUSSION

4.1. Humidity Effects and Optimization

Figure 1 shows time series of three units connected with different inlets. The microAeth with “regular” tubing exhibits large positive (e.g., 47.4 μg/m³ of BC at 13:33) and negative (e.g., −67.4 μg/m³ of BC at 13:32) excursions or spikes in the BC data while the microAeth units with diffusion driers showed clean flat-topped step functions with time, demonstrating the efficacy of the diffusion drier inlets in ameliorating the impact of rapid changes in RH and T when the units and inlets are worn. Similar false excursions have been observed during rainy day sampling using a microAeth without a diffusion drier (Cai et al. in review) and using other optical instruments, such as nephelometer (Arnott et al. 2003; Fischer and Koshland 2007).

Rapid changes in RH and T can result in evaporation or condensation of moisture from the surface of the collecting medium (especially hydrophilic filters) and/or optical components. Differences in the RH and T impacts on the sensing channel versus the reference channel optical paths is the working model of the cause of the false BC spikes. In the current design of the microAeth, the spot on the filter seen by the reference detector does not have any active flow of incoming air and thus is not as quickly affected by changes in RH and T as the sensing channel detector. The reference channel has to wait for diffusion/advection of moisture. As shown in Figure 1, the BC reading of the unit with only the regular inlet tubing returned with time to normal levels (i.e., the relatively constant levels recorded by the units with the diffusion driers), suggesting the humidity conditions of the reference channel and sensing channel eventually equilibrate with environmental changes in RH and T; the amount of time it took ranged from 4 min (cycle OUT 4) to ∼12 min (cycle IN 2).

The effectiveness of the diffusion drier is partially explained by reducing incoming moisture and thus buffering the abrupt humidity change and lowering chances of condensation and false excursions in BC readings. Comparing the rate of change in the RH (Figure S1 in the online supplemental information), one observes that the long and short diffusion driers keep the rate of change <2%/min and <2.5%/min, respectively, whereas the regular inlet (no diffusion drier) has a rate of change >∼10%/min when switching location between inside the chamber and outdoor.

The thermal heating and buffering when wearing units on persons in a vest pocket is a key part of this silica gel-based diffusion drier inlet system. For example, if one does not wear the monitors but instead places them on an open tray and just walks in and out of air conditioned buildings carrying the tray, the units with the diffusion driers can still have large negative and positive excursions in BC (Cai et al. in review). This is consistent with the diffusion driers being able to cause differences (i.e., impacts on light transmission due to different humidity levels) between the sensing channel and reference channel when not worn on the body. Since indoor settings do not experience rapid changes in RH and T, units used as fixed-site indoor residential monitors do not need inlet heaters or diffusion driers.

4.2. Vibrational Effects

During the initial deployment of duplicate microAeth units when the units were almost brand new, the BC average was 2.0 ± 3.6 μg/m³ (standard deviation, SD) and the %Diff between two units was 1.0% ± 31% (SD) for 1-min BC concentration. Examination of the minute-by-minute data shows that individual events and BC sources (Figure 2a) could be clearly identified from time-activity information (outdoors, subway and bus riding and cooking). Excluding two 1-min outliers, that may have been caused by close proximity to a BC source(s), overall R (Pearson correlation coefficient) between two units was 0.94 (p < .001). The two units correlated to a greater degree at higher BC levels; as can be seen R increases from 0.73 (p < .001) to 0.96 (p < .001), when BC level (±SD) increases from 0.9 (±0.5) μg/m³ (suburban area) to 4.5 (±4.8) μg/m³ (urban area). The reduced correlation observed at lower BC levels is probably a combination of variation being smaller due to the lower BC levels, and the lower BC levels experiencing relatively greater impacts from random electronic noise of the units. In summary, when the units were relatively new, personal sampling provided robust data with good agreement between units used as personal samplers, as seen in Figure 2a.

FIG. 2 (a) Time series of duplicate microAeth personal sampling with BC sources when the microAeth units were relatively new. Note that the enhanced levels of BC during time in the subway are probably caused by the abraded steel wheels which creates high concentrations of steel dust and airborne black iron oxides (Chillrud et al. 2004), which has significant absorption at the wavelength used by the microAeth. Due to the lower sensitivity to magnetite (Yan et al. 2011), as far as we are aware, it is only in the special microenvironment of enclosed subway tunnels that these black iron oxides obtain concentrations that can be recorded by the microAeth. (b) Time series of duplicate microAeth personal sampling after 17 months of heavy use. About 45-min of scripted physical movements, including walking, jumping, and running, were made during the suburban sampling when BC levels were near the limit of detection (0.1 μg/m³).

However, after 17 months of heavy use, we found that microAeths were much more sensitive to vibration. Example data are shown in Figure 2b (unit 102 vs. unit 109) where large positive and negative spikes in the real time BC data were associated with time periods of physical movement, even in routine movement (e.g., subway and bus ride), which was not an issue when the instrument was new (Figure 2a). The 1-min %Diff between these two units were 176.0% ± 2444% and 35.6% ± 191% for the personal sampling in suburban area and urban area, respectively. The 1-min data for the duplicate units did not correlate significantly (p > .05) during the personal sampling in suburban area, mainly due to (a) the low BC concentration (about 0.1 μg/m³; note that the limit of detection for the microAeth monitor at 100 ml/min flow rate and 1 min integrations, based on 3× standard deviation of signal from filtered air is about 0.1 μg/m³) and (b) the periods of scripted physical activity/movement. However, personal duplicate sampling during periods of higher BC levels (0.9 ± 0.4 (SD) μg/m³) increased R to 0.94 (p < .001). The %Diff between the units were approximately one order of magnitude higher than these statistics when the units were relatively new. Over the whole time period, the data quality of personal sampling decreased and becomes “spiky,” compared to the previous testing, especially during time periods of low BC levels. But this “spikiness” of the logged data greatly decreased during periods when BC levels were high and/or the units were not worn (i.e., no external physical movement). The %Diff of 1-min data was 11.9% ± 12%, during night-time periods when the compliance monitoring indicated the units were not worn. The night-time data quality was more similar to data collected with these same two units during 24-h fixed site sampling when units were relatively new (data not shown), where the %Diff of 1-min data between two units (unit 102 vs. unit 109) was 2% ± 10% (SD).

The sensitivity of microAeth real time data to physical movements also was observed in the field data collected from the cohort study of children. In 17 of the 19 successful duplicate deployments in the children's study, the raw (i.e., uncorrected) 5-min BC data correlated significantly (p < .05) with each other. However, R between two units ranged from 0.19 to 0.95. Two of the 19 deployments with duplicate sampling even showed nonsignificant correlations between duplicate units for 5-min raw data. One of these deployments employed a unit producing 20 min of continuous readings with LED light source error flag logged in the data file; after excluding these 20 min of data, the correlation of the two units becomes significant but still low (R = 0.4). The other deployment with nonsignificant correlation employed a unit that appeared to be much more sensitive to the physical vibrations than the other unit. Comparing the time series of BC data and compliance actiwatch data indicated that most of positive and negative excursions of the real time BC data occurred when subjects were active (Figure S2).

Multiple reasons can explain the increased sensitivity to physical movement with time. First, inspection of the optical inserts of the microAeths showed that large pieces of particulate matter (PM) and fibers were being retained on the optical insert and especially within the sensing channel of the insert. These large particles and fibers were able to easily move since they were not permanently stuck to the insert. This “dirt” moving in or out of the light path of either the sensing channel or reference channel can cause large positive and/or negative excursions in BC data. As shown above, the impact of these physical movements is dependent upon the overall level of BC being measured as seen in the duplicate experiments that included time spent in suburban areas (lower mean BC having higher impact of noise on %Diff) and urban areas (higher mean BC and having smaller impact on %Diff). Similarly, for locations that have multihour average BC concentrations at relatively high concentrations (e.g., greater than 10 μg/m³) observed in cities with rapidly developing economies, we have observed no to little impact of physical movements on the real time BC levels during personal deployments (data not shown).

This problem could likely be ameliorated using size selective inlets to reduce the amount of large PM and fibers that can enter the system, as well as regular cleaning of the optical inserts and flow paths. Cleaning a unit takes about 20 min and is best done by taking out the optical insert in a nondusty environment such as a high-efficiency particulate air (HEPA) filtered laminar flow bench and using a combination of techniques including filtered, clean compressed air, damp lint-free cloths, and warm deionized water.

Another reason for the increase in sensitivity to physical movement is that the hardware components degrade with time. Potential components include the rotary pump, the springs that hold the filter ticket in a fixed position, the tightness of the moving parts (bearings, shafts), and other electrical components including the power supply. Evidence for this is from the observation that the cleaning of the optical inserts and flow paths reduced the frequency of obtaining spiky data related to vibration but could not completely eliminate it (data not shown). Given that cleaning of the units was not sufficient, the microAeth units were sent to AethLabs for major maintenance and overhaul, including pump replacement, flow calibration and firmware upgrade. This overhaul resulted in major improvement in data quality (Figure S3).

FIG. 3 (a) Time series of ΔRef and ΔSen patterns and BC readings corresponding to stationary behavior of microAeth. (b) Time series of ΔRef and ΔSen patterns and BC readings corresponding to 10-min hand-shaken (23:05–23:15).

4.3. Development of Data Cleaning Algorithms

Given that cleaning of the optical flow paths was not able to eliminate vibrational impacts, a post-hoc data cleaning method was developed by comparing the patterns of Sensing and Reference signals under normal conditions and vibration for identifying the suspect data associated with vibration. As shown in Equation (4) (above), the BC measurement from the microAeth depends on both Δln(Sen)_(T2–T1) and Δln(Ref)_(T2–T1) and erroneous changes to the Sen and/or Ref signals at either T2 or T1 time points can lead to a false BC reading. Given Δln(Sen) and Δln(Ref) are normally very small values (absolute value <∼10⁻³) and not convenient for comparison, ΔSen and ΔRef are used instead in this article, which are the difference between two consecutive readings in the sensing signals and reference signals, respectively (i.e., Equations (1) and (2)).

Vibration-associated excursions were characterized by comparing the values of ΔRef and ΔSen in one unit that was run in a stationary manner (Figure 3a) to the values for the same time period for a second monitor that was shaken by hand (Figure 3b). In the stationary monitor, ΔRef wanders around zero, ΔSen tends to track changes in ΔRef up and down, and ΔSen is typically less than ΔRef and negative due to the increased attenuation by newly deposited BC (Figure 3a). In comparison, physical vibration such as shaking a monitor by hand can lead to up to three types of changes in the values of ΔRef and ΔSen, as shown in Figure 3b: (a) the magnitude of |ΔRef| (absolute value) during the time period of handshaking was larger than |ΔRef| values obtained for the stationary monitor, (b) sometimes ΔRef can be even more negative than ΔSen, and (c) ΔSen can have unrealistically high negative and/or positive values. Hence, looking at the changes in the reference and sensing signals as a function of time is very useful for identifying such events affected by rapid change either due to physical vibrations or changes in relative humidity (RH) and temperature (T) as well. However, type “c” events should be evaluated carefully to avoid flagging or removing valid data since very large negative ΔSen may be due to true filter loading during periods of high concentrations of airborne BC.

Based on the environmental chamber experiment and the vibration experiments, false BC data due to rapid changes in T and RH is discernable from false data due to vibration by looking at the pattern of ΔRef and ΔSen corresponding to known events (e.g., Figures 3a and b vs. Figure 4). For the false events related to rapid changes in RH and/or T, one sees large ΔRef and ΔSen changes that gradually and continuously increases or decreases together, from negative to positive and vice versa and this pattern can last for relatively longer time periods. This is mainly due to fact that there is no active flow passing through the reference channel. Thus, the reference channel takes relatively longer to catch up with the environmental changes in RH and T than the sensing channel, which does have active air flow. In comparison, vibration associated events are associated with more abrupt changes in ΔRef and ΔSens which can be in the opposite direction and which normally have a much larger amplitude than RH/T events.

FIG. 4 Time series of ΔRef and ΔSen patterns of the microAeth with only regular inlet tubing during the environmental chamber experiment shown in Figure 1. The missing data (15 min) was due to time spent downloading this unit's data as an initial check on the BC levels in the chamber to check on whether the single filter could be predicted to last the entire experiment.

On the basis of the experiments discussed earlier, two algorithms were suggested for automatically identifying and flagging problematic data. The first algorithm is based on |ΔRef|, which is supposed to vary around zero since no BC loading occurs on the reference area, and assigned a threshold value A (Equation (3)). The second algorithm is based on whether the ΔSen value, which typically is less than ΔRef and negative due to BC loading, was larger than ΔRef by a certain threshold (B) (Equation (4)). Data points are flagged if the thresholds of either Equation (3) and/or Equation (4) are met. In addition, the data point immediately after this suspect point is also flagged because the BC calculation of this data point relies upon the raw signals from the prior flagged data point to calculate the change in BC loading.

Threshold setting can be a double-edged sword; setting the threshold too low can lead to flagging some valid data points while setting it too high can lead to missing some data points that should be flagged. To find the optimum values, we first identified those problematic data points in a 24 h BC real-time data profile by comparing them with the profile collected by a more reliable unit run side-by-side, then we counted the number of points that were missed and wrongly flagged using different threshold setting. For example, when a unit was run at a 50 ml/min flow rate and a 5 min time base setting, values for Threshold A between 630 and 750 resulted in a good balance between over- and under-correction and with errors being less than 5% of total measurements (Figure S4). This was done for all units and similar thresholds were found. To be conservative, we used a value of 750 across all units. Threshold A values change when units are run with different setup parameters; for example, compared to the values above for 50 ml/min flow rate and 5 min time base, we found the best value for Threshold A is 175 when the unit was run at 100 ml/min flow rates and a 60 s time base. Threshold B, which is less affected by the flow rate, was set at 75 across all the units, independent of settings. These threshold values should be reviewed and potentially revised on a regular basis for each monitor or after major maintenance repairs. Using threshold values of 750 for Equation (3) and 75 for Equation (4), our observation is that more than 95% of vibration and/or humidity events can be identified.

Note that Equations (3) and (4) are unable to flag some BC readings that are affected by moving particles/fibers in the sensing channel when they produce large negative ΔSen but small ΔRef, which is a similar pattern with real BC loading. Furthermore, some BC readings that are affected by an unstable pump or light source (i.e., device is working improperly) without significant change in reference and sensing signals are not caught by the algorithms proposed here. However, these additional hardware type failures are caught by status error codes recorded in the data file (see SI Section IV for more detail). Our final QA/QC procedures apply flags to any reading with status error codes larger than a value of 1 and with readings flagged by the methods associated with Equations (3) and (4). In order to provide an estimate of how frequent these events are, we can look at the children's study data where a total of 124,765 readings were made; of these the data cleaning methods based on Equations (3) and (4) identify 14% of 124,765 readings as having issues and the data status flags picked up an additional 1% of the readings for a total of 15% of the readings were flagged (Table S1).

FIG. 5 The absolute value of the mean %difference between duplicate monitors for 19 duplicate deployments within the children's cohort, presented for different averaging times of the 5 min data. The absolute value of the mean is used to simplify the figure.

4.4. Validation of Data Cleaning Algorithms

Though there may be no need to clean data if one's research is only focused on longer term mean BC data (see below), for those studies interested in using the real time BC data, there are three options for what one can do with flagged real time data. The first option is just to remove the flagged data. Second is to impute the values of the flagged data by interpolating between the average of the readings before and after the suspect data. Third is to use the change in attenuation from before and after the flagged data to calculate the amount of BC that was deposited during the missing time interval.

All three data cleaning options improve reproducibility substantially, as can be seen by examining the absolute value of the mean %Diff between duplicate data from the children's cohort (N = 19), averaged over different time periods (Figure 5). For the data at 5 min averaging times, the |mean %Diff| = 26% ± 83% (SD) for the raw data versus 5% ± 43% for Option 1, 6% ± 39% for Option 2, and 10% ± 75% for Option 3. The |mean %diff| statistic for the raw data shows improvement with increased averaging time, suggesting that the longer time periods allows the false positive excursions to eventually be balanced by false negative excursions. For example, by 60 min, the absolute value of the mean %Diff of the raw data reaches ∼10%; while for the three data cleaning options, it was less than 5%. For 240 min averages, the raw data and three data cleaning options had very similar |mean %Diff| (≤3%). As such it appears that one can avoid data cleaning efforts if one is only interested in averages of BC longer than 4 h. Similar conclusions about averaging times were made from looking at the mean Pearson's correlation coefficient for the duplicate deployments made within the children's cohort study (Figure S6) and from looking at duplicate sampling experiments investigating the affects of humidity when comparing units with and without the diffusion drier inlet (Cai et al. in review), that is, the diffusion drier was not needed if one is only interested in BC levels representing averaging times longer than an hour.

Although the first data cleaning option appears to work well, deleting problematic data may result in data gaps long enough to miss some exposure events. The latter two options allow one to fill in these data gaps with imputed values. Considering that ambient BC concentration is a continuous variable, the second option is reasonable to use for two BC readings before and after the suspected data to replace the data gap. In concept, the third method can be more accurate, especially if very different concentrations of BC are occurring during the time period of missing flagged data as compared to before or after. However, Options 1 and 2 clearly showed a better statistical result than Option 3, both from the points of view that option 3 has higher |mean %diff| and larger standard deviations on the mean (Table S1).

Mean 24 h raw BC data agreed very well to the 24 h average of cleaned data (e.g., for option 1 above of processing the flagged data, R = 0.95, slope (±standard error, SE) = 0.98 ± 0.02, N = 330, p < .001; we excluded six of 336 sampling events in this comparison, four of them when the unit got wet during the sampling (liquid water was found inside of the unit after sampling) and another two deployments that had an extreme number of flagged data (>50% original data points).

5. CONCLUSIONS

When using the monitor as a personal or mobile monitor, then rapid changes in RH and T can lead to large positive and negative excursions in the real-time BC data. However, these impacts can be greatly ameliorated by the combination of wearing the device to buffer temperature changes and using a diffusion drier in the inlet. After significant and sustained use of the microAeth monitors, soiling combined with physical movement can result in units recording large positive or negative spikes in BC and having poor reproducibility. Though these issues may be ignored if one is only interested in using BC averages for periods ≥4 h or when overall BC level is very high (e.g., multihour mean >10 μg/m³), data cleaning of identified suspect events allows more confidence in the interpretation of the real-time personal monitoring data. Procedures to identify suspect data were developed and approaches to impute values for flagged data were suggested. The data cleaning methods increased the quality of microAeth real-time data, evidenced by improved reproducibility and increased correlations between duplicate deployments. If our suggested protocols and recommendations are followed, the microAeth can be used with confidence for personal, mobile, and fixed site monitoring for epidemiological studies focused on sources of incomplete combustion.

Acknowledgments

This research was supported by National Institute of Environmental Health Sciences (NIEHS) grants (ES016110, ES015905, ES013163, ES009089, ESO13163, and ES008977). We also gratefully thank Jeffrey Blair from Aethlabs (Aethlabs, San Francisco, CA, USA) for meaningful discussions on hardware and software issues. This is Lamont-Doherty Earth Observatory (LDEO) contribution number 7703.

[Supplementary materials are available for this article. Go to the publisher's online edition of Aerosol Science and Technology to view the free supplementary files.]