**2. Methodology**

Three Arduino Uno R3 (\$14.99 USD each), FONA808 modules (Adafruit, \$49.95 USD each) and dust sensors (GP2Y1010AU0F \$15.23 USD each) were purchased and used as received without modification. The dust sensor is a nephelometric sensor illuminating sample with a near infrared LED with a wavelength of 860 nm and collecting light at 120 degrees from incident illumination as depicted in Figure 1A. The dust sensors were a ffixed to circuit boards with cyanoacrylate and wired into Arduino 'stackable shields' in house by laboratory sta ff. Stackable shields plug into the Arduino board directly by using header pins for both electrical and mechanical connection. The wiring connections used are illustrated in the circuit diagram of Figure 1B. The FONA808, Arduino Uno, and dust sensors were mounted into plastic project boxes as depicted in Figure 1C. For user comfort, a lanyard was attached to each box. The Arduino Uno was used to control FONA808 modules and dust sensors. FONA808 communication modules were used for reporting the particulate matter concentration to a Thingspeak.com channel. Technical details of this process [40]. The data was collected in real time and

logged to Thingspeak.com until downloaded for analysis. All the calibrations were performed with wireless communication with 9V power supply powering the dust sensor modules.

**Figure 1.** (**A**) Photo and drawing of Sharp GP2Y1010AU0F dust sensor. (**B**) Wiring schematic for the dust sensor and FONA module to the Arduino (C) Photograph of the portable project box containing the dust sensor, FONA module and Arduino. The project box is 12 cm × 9 cm × 6.5 cm and has a mass of 250 g.

The Sharp dust sensors produce an analog voltage as output signal. To convert the analog signal to a meaningful PM concentration, the dust sensor must be calibrated against references before making meaningful measurements. A commercial device, the reference aerosol monitor (RAM-1) from Monitoring Instruments for the Environment (MIE, Inc., Billerica, MA, USA) was chosen as a reference method to calibrate each sensor. The RAM-1 actively samples aerosol using a pump and estimates concentration using the light scattering principle. The RAM-1 analysis is not an EPA equivalent method for PM2.5. We have performed four trials in which the RAM-1 sensor was used to compare the indicated concentration of a sodium chloride test aerosol within a chamber with gravimetric determination of aerosol mass concentration. A 37 mm quartz filter was used to collect a sodium chloride test aerosol at 15 SLPM using a mass flow controller and vacuum pump. Gravimetric measurements of (average) concentration over the sampling period could be obtained and compared directly with the RAM-1 values. For these 4 trials we find that the RAM reading is on average 107% (Std Dev. = 67%) of the gravimetric result with n = 4.

For calibration, all Sharp dust sensors were placed inside a 3.6 m<sup>3</sup> volume chamber lined with a fluoropolymer (FEP). Each dust sensor was individually calibrated in the 3.6 m<sup>3</sup> chamber. A single jet atomizer (TSI 9302) generated sodium chloride particles into the chamber with an atomizer pressure of 20 psi. The flow of particles entering the chamber was 5.7 L min−1. The jet generated polydisperse sodium chloride particles into the chamber. After approximately 2 min, su fficient aerosol was produced and the atomizer jet was shut o ff for the remainder of the experiment. The relative humidity inside the chamber under these conditions was measured to be < 20%, indicating presence of a dry aerosol. Experimental runs at high relative humidity were not attempted, and performance under these conditions not explored in this study. The calibration data was then collected for 4–5 h as the initially high particle concentration dissipated due to impaction on the chamber walls. For each point in time, the individual sensors were compared to the indicated concentration of the reference aerosol instrument (RAM) placed inside the chamber. Each dust sensor was calibrated individually, and the process was repeated for each sensor. All of the data from the dust sensors were uploaded to the ThingSpeak channel in real time using the FONA808 module.

#### **3. Results and Discussion**

#### *3.1. Optimizing Delay Time*

The Sharp dust sensors operate by turning on a near-IR LED, waiting a user-specified delay period, and then sampling the scattered light signal prior to cycling the LED o ff again. The entire measurement cycle can be completed in under a millisecond and this measurement cycle can repeat itself many times. Indeed, signals can be averaged to improve limits of detection (see section below). The LED is turned on by a HI/LO (5 - 0 Volt) transition at a digital pin on the microcontroller, and a delay prior to signal acquisition is then initiated. While Sharp recommends a delay time of 280 μsec prior to collection of data in product literature, we have systematically studied the e ffect of delay time on sensitivity (slope of calibration line). Figure 2 reports the results of this study. As observed, measurements sugges<sup>t</sup> that a 220 microsecond delay after the LED trigger allowed optimal sensitivity to be achieved for the individual sensors tested in this study. The cause of the shift in optimum delay time from the manufacturer's recommendation is not known. However, di fferences in stray capacitance between our apparatus and the manufacturer's test bed may be the cause. The result suggests end users may wish to optimize delay time for their own application. Regardless of the cause, the delay time has been optimized to provide maximum sensitivity for this study.

#### *3.2. Sensor Calibration Results*

Figure 3 illustrates plots of dust sensor signals vs. indicated PM mass concentration for the three dust sensors used in this study. In these experiments, signal from a commercial PM mass concentration monitor (RAM) was used as the reference/accepted concentration. Our analysis is limited in we assume no error, uncertainty or imprecision is present in the RAM data stream. For all sensors, there was a linear and positive correlation (R<sup>2</sup> > 0.92) observed between the concentration (μg m<sup>−</sup>3) and dust sensor signal to 500 μg m<sup>−</sup>3. A linear-least squares best-fit line was added to each dataset and used to determine the slope and intercept for each sensor. The slopes were 0.491, 0.446, and 0.506 digital counts per μg m<sup>−</sup><sup>3</sup> for sensors 1, 2 and 3. Since the Arduino analog acquisition is a 10 bit device operating over a full span of 1024 steps from 0 to 5 V, a digital step corresponds to 4.88 mV/count. Therefore, our sensors were observed to produce between 0.217 and 0.247 V per 100 μg/m<sup>3</sup> or roughly half of the signal specified by the product data sheet for the Sharp dust sensor which specifies 0.5 V signal per 100 μg/m<sup>3</sup> of PM mass concentration [41]. The exact cause of this discrepancy in sensitivity remains unclear. However, we note that Liu et al. [38] previously observed sodium chloride aerosol produced a much lower signal compared to other compositions. Varying the size distribution or particle shape of the test aerosol will lead to di fferences in di fferential scattering cross-section at 120 deg. and consequently, the observed slope of calibration lines. Emerging research on low-cost dust sensors [36,38] has provided

clear evidence that the light scattering method can be subject to large biases when projecting mass concentrations when calibrations were conducted against aerosol with different properties (refractive index, shape, size, composition) than the ambient type encountered at a particular location or for a particular application. As such, this could be responsible for the low sensor responses we report.

**Figure 2.** Plot of Sharp GP2Y1010AU0 dust sensor sensitivity vs. delay time prior to signal acquisition. The sensitivity is the slope of the best-fit calibration line when signal was plotted vs. PM mass concentration (μg/m3). The optimal delay of 220 microseconds was employed for subsequent measurements. Error bars represent ±1σ of replicate trials.

**Figure 3.** Calibration plots for sensors 1–3. Observed signal from the Sharp GP2Y1010AU0 dust sensors vs. PM concentration indicated by the RAM reference method for laboratory generated polydisperse NaCl aerosol. Observed signal is the digital count assigned by the Arduino board during analog-to-digital conversion. Differences were found in slope and intercept between the sensors tested, and R<sup>2</sup> > 0.92 observed. Relative humidity within the 3.6 m<sup>3</sup> chamber during calibration was < 20%.

#### *3.3. E*ff*ect of Signal Averaging*

Since the data acquisition cycle for a single pulse of the LED is very short (ms time), averaging many pulses from the LED is an attractive approach to enhancing detection limits for the Sharp dust sensors. Code can easily be added to the Arduino sketch to accomplish signal averaging in the onboard memory of the microprocessor on the fly. Only the averaged result is then reported in the data stream. A study was carried out in which an air blank was analyzed, and the number of LED pulses averaged together were 100, 200, 500, 1000, 1500, 2000, 5000, and 10,000 individual pulses. Data points were reported for each average and the standard deviation (σ) of replicate averages computed. Then, the limit of detection (L.O.D.) for each sensor was computed by using the 3σ standard. As observed in the black trace of Figure 4, as the number of LED pulses averaged increased to roughly n = 2000, the measurement L.O.D. decreased an order of magnitude for all the sensors tested. Interestingly, averaging approx. 2000 pulses only marginally a ffects the total time required to report a single data point (shown in blue data series) because the wireless communication protocol alone requires nearly a minute on average to complete its reporting cycle. Figure 4 also demonstrates that if additional LED pulses > 2000 were averaged together, there were only marginal additional gains in limit of detection achieved. However, the time required for data acquisition when n > 2000 began to linearly increase with the number of averages since wireless communication was no longer the rate limiting step under these conditions. Consequently, 2000 averages were used in subsequent monitoring activities. Under these conditions, the standard deviation and limit of detection for the three sensors were between σ = 3–6 μg/m<sup>3</sup> and L.O.D. = 9–18 μg/m3, respectively. These results indicate that Sharp dust sensors and microprocessor mediated signal averaging can be used to track PM pollution within environments where substantial particle pollution is expected. This could include routine workplace monitoring or personal exposure monitoring for citizens living in urban centers where levels regularly exceed 25–30 μg/m3. However, it should be noted that the method we describe herein is not sensitive enough for monitoring in all environments (see below). Further improvements in precision, limits of detection, and sensor accuracy are still required to bring ambient portable sensing to its full potential.

#### *3.4. Precision of Dust Sensors*

Figure 5A reports a histogram of observed percent di fference between individual Sharp dust sensor measurements and the reference method (RAM) measurements for test sodium chloride aerosol within a laboratory chamber. Figure 5 includes data points from all 3 Sharp dust sensors. Here, the percent di fference was computed as (Sharp sensor measurement–reference measurement)/reference measurement and the result expressed as parts-per-hundred relative di fference. As observed in Figure 5A, results were observed to follow a normal distribution. A Gaussian fit to the data indicated a standard deviation of σ = 30.8 for the entire dataset. This result suggests, considering all data, Sharp sensors report concentrations within 30.8% of the reference value about 68% of the time. The standard deviation also allows end users to define confidence intervals for their data points. In Figure 5B, a plot of percent di fference between Sharp dust sensor and the reference measurement is plotted vs. the indicated PM mass concentration for the reference method. This plot indicates that the largest relative discrepancies between dust sensor indicated mass and reference monitor mass occurred when PM mass concentration was < 40 μg m<sup>−</sup>3. Above this mass loading, relative di fference between the dust sensor and reference monitor was very frequently < 20%. Considering only data collected when PM > 40 μg m<sup>−</sup>3, the average percent di fference between a Sharp dust sensor and reference measurement for single point comparison was 5.8% and the median absolute percent di fference was 15.2%. This result indicates that unmodified Sharp dust sensors o ffer best precision when PM mass loadings are relatively high (e.g., PM > 40 μg m<sup>−</sup>3).

#### *3.5. Monitoring Experiments*

The ultimate goal of this work is to improve understanding of human exposure to PM pollution through creating and implementation of a network of multiplexed, portable PM sensors. In this manuscript, we report a limited scope proof-of-concept application by having laboratory personnel carry the three dust sensors around Lubbock, TX. The sta ff engaged in normal life activities during sampling in an e ffort to demonstrate that data from multiple sensors can simultaneously be streamed to a Thingspeak channel, while the end user can access information about his/her own exposure in nearly real-time via a public Thingspeak web link. Figure 6 reports example time-series monitoring data that end users encountered during the e ffort. As observed, spikes in PM mass loadings were encountered by the users during random life events such as walking near an individual smoking cigarettes, cleaning an apartment space, burning incense at home, or cooking. It is often di fficult to accurately parameterize exposure during such casual life experiences in models of human exposure to PM pollution, and portable PM sensors have grea<sup>t</sup> promise to improve quantification of PM for such circumstances. In an effort to improve understanding of human exposure in different environments we have performed measurements of PM mass concentration at a variety of locations and report a summary of these measurements in Table 1 of this manuscript. While a wide range of PM levels were encountered across locations (and even at single locations through time), the reported levels begin to improve understanding of typical levels of human exposure in various environments. For instance, results have suggested that human exposure to PM is often quite high near any cooking operation, or at some restaurants. The use of low-cost, portable sensors to uncover and document such knowledge can be used to develop exposure control strategies for workers in these environments.

**Figure 4.** The effect of number of LED pulses averaged (N) on the limit of detection (black squares) for sensors 1–3. The second y-axis (blue) presents the seconds required to acquire and report data for the specified number of averages.

**Figure 5.** (**A**) Histogram of percent difference between individual Sharp GP2Y1010AU0 dust sensor measurements and the reference method measurements (RAM-1) for test sodium chloride aerosol within a chamber. A Gaussian fit to the data indicated σ = 30.8 for the entire comparison dataset. (**B**) Plot of percent difference between Sharp dust sensor and the reference measurement plotted vs. the indicated PM mass concentration for the reference method.

**Figure 6.** Sample monitoring results for end users during the trial implementation. Real-time personal exposure monitoring allows improved understanding of the times, locations, and extent of high exposure events. The sensors were in the same room adjacent to the end user during the events.

Another goal of future research is to begin documenting the temporal and spatial behavior of aerosol levels at specific locations. Figure 7 presents a map of PM levels encountered by end users as they carried the portable sensors with them through their daily routines. Since the platform developed transmits GPS coordinates along with PM levels, data from multiple sensors can be combined and plotted. Such plots can ultimately be used to better understand the spatial dynamics of particle pollution within a city, and comprehensive datasets can better educate the public in regard to differences in pollution exposure between regions of a city or neighborhoods. However, the results presented in this manuscript do not represent such an achievement. Understanding temporal and spatial averages for PM concentrations at specific locations is a complex task that can only be accomplished through many more measurements than what we report within this paper.



**Note:** Table 1 data is not representative of exposure in all specified microenvironments. Human exposure to PM is highly circumstantial based upon an individual's specific experience.

**Figure 7.** Maps of Lubbock, TX with user sampled PM single-point data superimposed. Mobile sensors transmitted GPS coordinates and dust sensor PM measurements to a webserver, and maps were created from data. Map in (**A**) is approx. 40 km wide by 28 km high. Map in (**B**) is zoomed to illustrate specific locations sampled by users during their experiences. Data point colors describe an individual observation PM mass concentration in μg m<sup>−</sup><sup>3</sup> and are not meant to represent trends or averages.
