The experimental set up for performance tests is shown in Figure 2-A. The test chamber consisted of a mixing zone (0.64 m × 0.64 m × 0.66 m) and a sampling zone (0.53 m × 0.64 m × 0.66 m), divided by a perforated plate positioned in the middle of the test chamber. The perforated plate contained 600 evenly spaced holes, each with a diameter of 0.6 cm. The perforated plate provided a homogenous airflow with no dead zones inside the sampling zone. The aerosol from the generation systems was diluted by clean air from two HEPA filters (0.25 m³/min) and mixed with a small fan in the mixing zone. The wind speed in the sampling zone was 0.01 m/s, resulting in a Reynold number of 400 (laminar flow). Three of each low-cost sensor (DC1700, Sharp DN, and Sharp GP) and one pDR-1500 operated with an inlet cyclone (cut-off diameter of 10 μm) were positioned in the sampling zone. The pDR-1500 was operated in active mode with a 37-mm glass microfiber filter (934-AH, Whatman, USA) at the outlet. The high-cost reference instruments were outside the test chamber, with direct sampling from the sampling zone.

Five polydispersed aerosols were generated using four different aerosol generation systems as depicted in Figure 2-B. Salt is a common environmental aerosol and a common test aerosol used to evaluate aerosol instruments. Arizona road dust is representative of a coarse mineral dust (Curtis et al. 2008) commonly found in environment and occupational settings and commonly used to calibrate direct-reading instruments. Diesel fumes are common in environmental and occupational settings, and welding fume is a critical occupational hazard. To achieve two aerosols of different size with the same refractive index, salt aerosols were generated using a Collison-type nebulizer (Airlife, company, USA) using two salt solutions (mass fractions of 0.9% and 5%) (Figure 2-B(I)). This aerosol was diluted with clean air and mixed in a chamber (0.1 m³) to achieve desired concentrations. We used a fluidized bed aerosol generator (3400A, TSI, USA) to aerosolize Arizona road dust (Fine Grade, Part No. 1543094., Powder Technology INC., Arden Hills, MN) with the concentration adjusted by controlling the feed rate of the dust entering the fluidized bed (Figure 2-B(II)). Diesel fumes were produced as exhaust from a diesel electric generator (DG6LE, Red Hawk Equipment, USA) with a valve used to waste fume and control concentrations (Figure 2-B(III). Welding fumes were generated with a welding system (0.03 inch Flux-Corded MIG Wire, Campbell Hausfeld, USA) operated inside a sandblast cabinet (Item 62454, Central Pneumatic, Byron Center, USA) [Figure 2-B(IV)]. To control concentrations, varying amounts of HEPA filtered air were used to push the fume from the cabinet to the sampling chamber.

The concentration of aerosols in the test chamber for each experiment fell into various ranges dependent on three factors: 1) measureable range of the DC1700 (0 – 231 particles/cm³); 2) maximum aerosol concentration of our experimental set up and equipment; and 3) concentration levels that range from 0 – 6500 μg/m³. Although concentrations were lower than OSHAs occupational exposure limit for particles not otherwise specified (15,000 μg/m³), these concentrations are relevant to the needs of practicing industrial hygienists, who often take action to control contaminants when concentrations reach one-tenth the limit. Steady-state concentrations of test aerosols were maintained at different levels. Aerosol size distribution varied by particle type, but was approximately the same for each concentration level of the same aerosol type, except for diesel fume (Figure S2 in online supplemental information). For each level, the number concentration by size was measured with the SMPS three times after reaching steady-state concentration. The APS was set to record particles number concentration by size every minute throughout the experiment. Prior to starting experiments, the air in the chamber was confirmed to be clean with the pDR-1500 (0 μg/m³) and the CPC-3007 (0 particles/cm³).

As a measure of sensor precision, we calculated the coefficient of variation (CV) of the raw sensor output (number concentration for DC1700 and voltage for the Sharp sensors) and after conversion to mass concentration (described below). For each minute, CV was calculated as (NIOSH 2012): (2)CV=σμ where σ is the standard deviation and μ is the mean of the measurements from the three replicate sensors of the same type. The average of one-minute CVs obtained over all concentrations for a given aerosol were reported for each sensor. EPAs acceptable CV values for test instruments are up to 10% (EPA 2016). NIOSH does not have a specific acceptable value for CV.

We used the pDR-1500 to develop equations to convert raw output from the low-cost sensors to mass concentration. The pDR-1500 was selected as a small and portable device within the budget of many corporate industrial hygiene programs that may then potentially be used to develop correction models in the field. For each aerosol type, the mass concentrations from the single pDR-1500 were averaged to match the one-minute DC1700 measurements. These values were then adjusted by multiplying by the mass concentration measured gravimetrically with the glass microfiber filter internal to the pDR-1500 and dividing by the average of unadjusted mass concentration from pDR-1500 over the entire experiment. Different filters were used for each aerosol type measured. Linear regression was used to determine the best-fit line between the 1-minute paired number concentrations of the DC1700 (small bin or total particles) and the corrected pDR-1500 mass concentrations. For the Sharp sensors, linear regression was used between the 1-minute paired Sharp voltages and the corrected pDR-1500 mass concentrations.

We evaluated the mass concentrations estimated for the low-cost sensors with reference to gravimetrically-adjusted mass concentration measured with the SMPS and APS following procedures specified by EPA for FEMs (40 CFR Part 53, Subpart C and 40 CFR Part 58) and NIOSH for evaluating direct-reading gas instruments (NIOSH 2012). For each three-minute SMPS measurement, the number concentration by electrical equivalent mobility diameter from the SMPS and the number concentration by aerodynamic diameter from the APS were converted to mass concentration by volume equivalent diameter, using the particle density and shape factor provided in Table S1 (online supplemental information). The reference mass concentration was calculated as SMPS/APS mass concentrations summed over all sizes multipled by the mass concentration measured gravimetrically with the glass microfiber filter internal to the pDR-1500 and divided by the mean of unadjusted SMPS/APS mass concentrations over the entire experiment. Three-minute averages of mass concentrations from the low-cost sensors and the pDR-1500 were calculated to correspond in time with the reference concentrations. For each aerosol and sensor, Pearson coefficient (r), coefficient of determination (R²), slope, and intercept between the sensor and reference mass concentration were determined using linear regression.

The bias (B) was calculated as (EPA 2016): (3)B=1n∑yi-xixi where y is the estimated mass concentrations for low-cost sensor from the regression models, x is the gravimetrically filter corrected SMPS and APS mass concentrations, and n is the number of data pairs.

These parameters were compared to acceptance criteria from EPA and NIOSH. For EPA, the linear regression between measurements made with a candidate and reference instrument must have a slope of 1 ± 0.1, a y-intercept of 0 ± 5 μg/m³, and an r ≥ 0.97 (40 CFR Part 53, Subpart C, Table C-4). EPA also specified that percent bias should be within ±10% (40 CFR Part 58). For NIOSH, candidate and reference instruments must exhibit a slope of 1 ± 0.1 and a percent bias of ± 10%. NIOSH does not have criteria for the y-intercept (NIOSH 2012).

Table 2 summarizes the precision expressed as the coefficient of variation (CV) calculated for raw data and converted mass concentration data by sensor and aerosol type. Scatter plots of raw sensor response relative to reference mass concentrations (i.e., filter-corrected SMPS and APS data) are shown for the DC1700 in Figure 4, and for the Sharp DN and Sharp GP in Figure 5. The x-axis error bars in these figures (the standard deviation of reference measurements) indicate the variability of mass concentration at each steady state. The y-axis error bars (the standard deviation within sensor type) indicate a combination of within-sensor precision and concentration variability. We displayed Figure 5 with log-log axes to allow all data to be visible (particularly for the Sharp GP) and Figure 4 in the same way for consistency. Based on the raw measurements, the DC1700 exhibited the best precision of the three sensors with the lowest CV (2.2–14% for the small bin; and 5–15% for the large bin) and smallest error bars in the response curves (Figure 4). For the Sharp sensors, the CV was high for most cases, varying from 17–30% for the Sharp DN and 2–51% for the Sharp GP.

Sensor precision is an important factor for determining an approach to estimate mass concentration from their response. The good precision of the DC1700, presumably stemming from the fact that the factory calibrates each sensor (Northcross et al. 2013), suggests that all DC1700s can be treated similarly when estimating mass concentration. In contrast, the Sharp sensors are substantially less precise, indicating the need to treat each sensor independently when computing mass concentration. In regards to the Sharp GP sensor, Budde et al. (2012) correlated multiple Sharp sensors with a DustTrak and reported that each sensor has different measurements. The author attributes these differences to fact that the Sharp sensors are not factory calibrated to reproduce the same values. Wang et al. (2015) measured the precision of the Sharp GP sensor, and found that the standard deviation among devices increased with mass concentrations, concluding that the Sharp GP sensors have low precision. To our knowledge, no one has reported the precision of the DC1700.

For all sensors, aerosol type strongly influenced sensor response. The response relative to reference mass concentration for the DC1700 (Figure 4) was linear for low concentrations and deviates from linearity for particle concentrations greater than 106 particles/cm³. The change in response relative to change in mass concentration (slope) was greatest for the salt aerosol, with both salt solutions having similar slopes. The slopes for welding fume and Arizona road dust were similar but less than that for salt aerosol. The slope was the least for the diesel fume, dramatically different from that of the other aerosols. In contrast, the response of the Sharp sensors (Figure 5 and Figure 6) was linear for all aerosols, except for the diesel fume. The Sharp DN (Figure 5(A)) and Sharp GP (Figure 5(B)) response curves had similar correlations but with different magnitudes. The slope of the Sharp sensors was similar to that of the DC1700, except for the fact that the slope was lower for welding fume than diesel fume. The response of the Sharp GP voltage with the mass concentrations for Arizona road dust was inconsistent with the manufacturer’s published data (Sharp 2006).

The different responses for the low-cost sensors to different aerosol types is a clear indication that each sensor needs to be correlated independently with a reference instrument for different aerosol environments. Particle size distribution, refractive index, and shape affect particle light scattering. The mass median diameters (MMDs) and geometric standard deviations (GSDs) of each aerosol obtained from SMPS and APS data differed substantially (Table 3). The MMD for the 5% salt solution is higher than the 0.9% salt solution, which was expected. The MMD for the Arizona road dust was around 2.6 μm. We did not calculate MMDs and GSDs for the diesel fume and welding fume aerosols because the size distributions were multimodal (see Figure S2 and Figure S3 in online supplemental information, respectively). A complicating factor is that the particle size distribution of the diesel fume changed for different concentration levels (Figure S2 in online supplemental information). The refractive index for the salt aerosol, dust, diesel fume, and welding fumes are 1.544, 1.51, 1.465, and 1.8, respectively. Since both salt aerosols have nearly identical slopes but different MMDs, we can infer that the refractive index has a role in the response. Northcross et al. (2013) and Wang et al. (2015) both suggest that the refractive index plays a role in response variation for different aerosol types. We observed that the response of the DC1700 to diesel fumes was substantially different than that for other aerosols. This finding may relate to the fact that the refractive index of diesel fume has a high absorption component or that the size distribution shifted with particle concentration.

The finding that the response of the DC1700 for high concentrations (>106 particles/cm³) is non-linear was expected. The non-linear response above 106 particles/cm³ is consistent with coincidence as specified by the manufacturer. The performance of the DC1700 was analyzed for number concentrations below 106 particles/cm³. This number concentration is equivalent to a mass concentration of 500 μg/m³ for both the 0.9% and 5% salt solutions, 1200 μg/m³ for Arizona road dust, 6300 μg/m³ for diesel fume, and 2200 μg/m³ for welding fume (Figure 4). For the Sharp GP sensor, Wang et al. (2015) found similar linear response between the low-cost sensor and the SMPS for measuring NaCl particles. Steinle et al. (2014), Holtius et al. (2014), Klepeis et al. (2013), and Northcross et al. (2013) all observed a linear relationship for relatively low concentrations (<300 μg/m³), whereas Semple et al. (2013) found a non-linear relationship for concentrations that exceeded 1000 μg/m³.

Linear equations used to convert sensor raw output to gravimetrically adjusted mass concentration measured with the pDR-1500 are shown by aerosol type in Table 4. For the DC1700, we only fit data for number concentrations <106 particles/cm³ (small bin) to avoid issues with coincidence. Limiting the data to concentrations below 106 particles/cm³ resulted in good linear fit with an R² value ranging from 0.91 for welding fume to 0.99 for 5% salt solution. For the Sharp sensors, we determined a linear equation unique to each sensor and aerosol because of their low precision. These equations are not shown in this manuscript because they cannot be used for other Sharp sensors, even for the same aerosol type.

We calculated again the CV values for the low-cost sensors based on the mass concentrations, as shown in Table 2. Based on the estimated mass concentrations, all the low-cost sensors have a low CV value that ranged from 1–8%. The low CV values based on the estimated mass concentrations are an indication that any sensor can be used to estimate mass concentrations once calibrated for an aerosol type.

Good agreement was observed between mass concentrations measured with the pDR-1500 and reference mass concentrations (from SMPS APS data; Table 5). Values of r were all high and the slope values were close to unity (ideal). The coefficient of determination (R²) ranged from 0.98 to 0.99. Biases were low ranging from −1.1% for welding fume to −9.1% for diesel fume. These results suggest that the pDR-1500 is a good medium cost device that can be used to determine regression models for the low-cost sensors.

Table 6 summarize the evaluation of mass concentrations estimated with the DC1700 relative to those measured with the SMPS and APS. Values of r and R² were high (an ideal value is 1). However, slopes were closer to unity and biases were <10% for the salt and welding fume aerosols but exceeded 18% for the Arizona road dust and diesel fume aerosols. The DC1700 calculated mass concentrations with the reference mass concentrations are shown in Figure S4 (online supplemental information). The points in Figure S4 approach the one to one line. Northcross et al. (2013) also reported the linear regression between the mass concentrations estimated from the DC1700 to those from a reference instrument for three ambient sampling periods. In that work, the slope and R² values varied between 0.95–1.96 and 0.81–0.99, respectively. These higher slope and lower R² values compared to our work may be because their study was conducted in an uncontrolled ambient setting with many different particle types sampled simultaneously.

Table 7(A) and 7(B) summarize the evaluation of the Sharp DN and Sharp GP sensors, respectively. Both Sharp sensors behaved similarly. For a given aerosol, the values of slope and intercept were similar for the Sharp DN and Sharp GP sensors. For example, the Sharp DN for the 0.9% salt solution experiment measured a maximum of 2.9 volts, whereas the Sharp GP only measured 0.9 volts for the same experiment. Slopes for both sensors were near unity, ranging from 0.9 to 1.3. The coefficient of determination (R²) ranged from 0.95 to 0.99. Biases for the sharp sensors, ranging from −9.8% to 5.2%, were substantially lower than those observed for the DC1700 (Table 6). The fact that these values were similar for both Sharp sensors can be attributed to the fact that mass concentrations were estimated from correlations with the pDR-1500. The Sharp sensors calculated mass concentrations compared to the reference mass concentrations are shown in Figure S5 (online supplemental information). The points in Figure S5 approach the one to one line.

Bias in mass concentrations estimated with the DC1700 and reference instruments met the acceptance criterion for 0.9% salt solution, 5% salt solution, and the welding fume, but not for the other two aerosols. The correlation coefficient met EPA’s criterion for all the aerosols except diesel fume. The slopes for all aerosols (excluding dust and 5% salt solution) met acceptance criterion. However, no intercepts met the acceptance criterion. For the Sharp sensors, bias and correlation coefficients met acceptance criteria. Slopes met the criterion, except for 0.9% salt and Arizona road dust aerosols. Only the intercept for 5% salt solution met the EPA criterion.

There are three main limitations to this study. The detection efficiency of the DC1700 (Figure 3) was measured using two aerosol types (salt for particles smaller than 0.3 μm and an oleic acid for particles larger than 1.3 μm), which have different refractive indexes. For diesel and welding fume, assumed values for density and shape factor introduced uncertainties in reference mass concentrations from SMPS and APS data. Particles larger than 300 nm were assumed to have a constant shape factor, although Park et al. (2004) reported that shape factor increases with particle size for diesel fume and Kim et al. (2009) reported the same for welding fumes. We also assumed a constant density for diesel fume, although Park et al. (2004) report that density decreases with particles larger than 300 nm. Lastly, the difference in MMD between salt aerosols was smaller than anticipated, yielding little information on the effect of size on sensor response.