Mapping Occupational Hazards with a Multi-sensor Network in a Heavy-Vehicle Manufacturing Facility

Abstract

Due to their small size, low-power demands, and customizability, low-cost sensors can be deployed in collections that are spatially distributed in the environment, known as sensor networks. The literature contains examples of such networks in the ambient environment; this article describes the development and deployment of a 40-node multi-hazard network, constructed with low-cost sensors for particulate matter (SHARP GP2Y1010AU0F), carbon monoxide (Alphasense CO-B4), oxidizing gases (Alphasense OX-B421), and noise (developed in-house) in a heavy-vehicle manufacturing facility. Network nodes communicated wirelessly with a central database in order to record hazard measurements at 5-min intervals. Here, we report on the temporal and spatial measurements from the network, precision of network measurements, and accuracy of network measurements with respect to field reference instruments through 8 months of continuous deployment. During typical production periods, 1-h mean hazard levels ± standard deviation across all monitors for particulate matter (PM), carbon monoxide (CO), oxidizing gases (OX), and noise were 0.62 ± 0.2 mg m⁻³, 7 ± 2 ppm, 155 ± 58 ppb, and 82 ± 1 dBA, respectively. We observed clear diurnal and weekly temporal patterns for all hazards and daily, hazard-specific spatial patterns attributable to general manufacturing processes in the facility. Processes associated with the highest hazard levels were machining and welding (PM and noise), staging (CO), and manual and robotic welding (OX). Network sensors exhibited varying degrees of precision with 95% of measurements among three collocated nodes within 0.21 mg m⁻³ for PM, 0.4 ppm for CO, 9 ppb for OX, and 1 dBA for noise of each other. The median percent bias with reference to direct-reading instruments was 27%, 11%, 45%, and 1%, for PM, CO, OX, and noise, respectively. This study demonstrates the successful long-term deployment of a multi-hazard sensor network in an industrial manufacturing setting and illustrates the high temporal and spatial resolution of hazard data that sensor and monitor networks are capable of. We show that network-derived hazard measurements offer rich datasets to comprehensively assess occupational hazards. Our network sets the stage for the characterization of occupational exposures on the individual level with wireless sensor networks.

Introduction

Low-cost sensors have attracted the attention of environmental health scientists interested in measuring air pollution with a high degree of spatial and temporal resolution, despite the sensors’ lower accuracy, precision, sensitivity and specificity (Snyder et al., 2013; Piedrahita et al., 2014; Kumar et al., 2015; Masson et al., 2015; Lewis et al., 2016). Current sensor availability reflects regulatory and health priorities (Lewis and Edwards, 2016); for instance, many low-cost sensors are available for particulate matter (PM) (Jovašević-Stojanović et al., 2015; Sousan et al., 2016) and hazardous gases, such as carbon monoxide (CO), ozone (O₃), nitrogen dioxide (NO₂) and sulfur dioxide (SO₂) (Xiong and Compton, 2014). In the current study, a sensor network was developed to measure PM, CO, O₃, and noise, agents that are important for worker health in industrial manufacturing facilities. PM has well-established relationships with cardiopulmonary and respiratory diseases, lung cancer, inflammation, oxidizing stress, pulmonary infection, and lung function (Dockery, 1993; Pope et al., 1995; Pope and Dockery, 2006; Anderson et al., 2012). The permissible exposure limit (PEL) for respirable PM is 5 mg m⁻³ (OSHA, 1993). The health effects of CO at and below the PEL, which is equal to 50 ppm (OSHA, 1993a), include headache, dizziness, weakness, nausea and confusion (Raub et al., 2000). Ozone is a well-known oxidant and its inhalation causes inflammation, reduced lung function, DNA damage and increased symptoms and development of asthma (Lippmann, 1989; Bornholdt et al., 2002; Weschler, 2006; Kampa and Castanas, 2008). The PEL for O₃ is 100 ppb (OSHA, 1993). Occupational noise exposure induces hearing impairment, hypertension and annoyance and may be associated with biochemical effects, immune effects and changes in absentee rate and performance (Passchier-Vermeer and Passchier, 2000). The permissible exposure to noise for an 8-h work period is 90 dBA (OSHA, 1974).

The standard technique for quantifying PM in the occupational environment is gravimetric analysis, which requires filters, pumps, and an analytical balance (NIOSH, 2017), or a third party to weigh filters. Although a variety of strategies are employed to sample for hazardous gases in the workplace, such as detector tubes, whole air sampling, sorbent sampling, and direct-reading instruments (DRIs), each has advantages and disadvantages (Harper, 2004). Major disadvantages of these strategies include the need for trained professionals, equipment requirements, sample handling, high cost, and large size/weight (Harper, 2004). Additionally, for both PM and gases, a major drawback of these sampling strategies (except for DRIs) is that measurements are time integrated—commonly eight to 10 h—reflecting typical work shifts. Sensors can overcome some of these disadvantages of traditional methods because of their low cost, small size, high temporal resolution, portability, and low power consumption (Lee and Lee, 2001; Xiong and Compton, 2014). However, the drawbacks of sensors include the need for thorough laboratory/field calibration, lower accuracy, precision, sensitivity and specificity, cross-sensitivity with non-target species, and instability over time compared with traditional techniques (Mead et al., 2013; Lewis and Edwards, 2016; Lewis et al., 2016). Despite these challenges, sensors offer a complementary strategy to study human exposure to occupational and environmental hazards.

Recent advances in open software toolkits and the modularization and commoditization of microprocessor platforms have facilitated the development of customized wireless sensor networks applications. Sensor networks are collections of small, inexpensive devices distributed throughout an environment (Heidemann and Bulusu, 2001). A growing number of examples of environmental health sensor networks (Kumar et al., 2011; Ikram et al., 2012; Mead et al., 2013; Gao et al., 2015; Hasenfratz et al., 2015; Heimann et al., 2015; Moltchanov et al., 2015; Jiang et al., 2016; Jiao et al., 2016; English et al., 2017) offer potentially powerful tools for hazard mapping, a technique that displays measured hazard(s) throughout a facility or geographic area (O’Brien, 2003; Peters et al., 2006; Koehler and Volckens, 2011; Koehler and Peters, 2013). Hazard maps can be used to visually communicate risk (Koehler and Volckens, 2011), identify hazard sources (O’Brien, 2003; Evans et al., 2008), characterize the distribution of hazards in a facility or the environment (Peters et al., 2006; Evans et al., 2008; Ott et al., 2008), and inform hazard control strategies (O’Brien, 2003). However, although the number of sensor networks deployed in the ambient environment is growing quickly, they are rare in occupational environments and represent an opportunity to increase the characterization of hazards at low cost. While aerosol mapping in particular has been successful in a variety of occupational and ambient settings (O’Brien, 2003; Peters et al., 2006, 2012; Heitbrink et al., 2007; Evans et al., 2008; Ott et al., 2008; Liu and Hammond, 2010; Park et al., 2010; Vosburgh et al., 2011), there is great potential for mapping other hazards such as noise, vibration, radiation, gasses, and vapours (Koehler and Volckens, 2011).

Hazard mapping ideally uses frequent measurements at high spatial resolutions to reflect the spatial and temporal variability of the hazard (Evans et al., 2008). However, hazard mapping is traditionally conducted with a limited number of DRIs that are transported through time and space during surveys. This practice requires data interpolation because measurements likely fail to portray the temporal variability present (Koehler and Volckens, 2011; Lake et al., 2015), and in some cases temporal variability can be incorrectly interpreted as spatial variability (Ludwig et al., 2017). Sensor networks have the potential to avoid a major pitfall in hazard mapping by reducing errors due to data sparsity (or ‘completeness’) that arise from the inability to measure a hazard at all locations and times simultaneously (Koehler and Volckens, 2011; Lake et al., 2015). In addition to errors of completeness, DRIs are prone to other errors also, including poor accuracy or precision, lack of sensitivity and biases from interferences (Koehler and Volckens, 2011). Using sensors to map hazards instead of DRIs will likely result in larger errors of these types.

In this study, we report on the deployment of a sensor network in a heavy-vehicle manufacturing facility, capable of measuring multiple agents of occupational interest including PM, hazardous gases and noise simultaneously and in real time, for a study period of 8 months. Although low-cost sensor networks have previously been used in the general environment, our network is the first of which we are aware that has been deployed in the industrial setting.

Methods

Multi-hazard monitor

We designed and constructed 50 multi-hazard monitors and deployed 40 in this network (Thomas et al., 2018), holding 10 monitors in reserve to replace monitors in the facility in the event of malfunction or failure. Briefly, each monitor was equipped with a dust sensor to measure PM (GP2Y1010AU0F, SHARP Electronics, Osaka, Japan); an oxidizing gas (O₃ + NO₂) sensor (OX-B421, Alphasense Ltd., Essex, UK); a CO sensor (CO-B4, Alphasense Ltd., Essex, UK); a custom sound pressure level (SPL) sensor to measure noise (Hallett et al., 2018); and a temperature and relative humidity sensor (AM2302, Adafruit, New York, NY, USA). A microcontroller (Seeeduino Cloud, Seeed Technology Co., Ltd., Guangdong, PR China) was programmed to read the electric signals from each sensor every 2 s and then average the signals and wirelessly transmit the averaged data to a central database approximately every 5 min.

Sensor calibration

Due to constraints at the facility, individual calibration of each PM sensor in the network was not feasible. Instead, the PM sensors used in the study were selected for inter-sensor agreement in the laboratory, and then underwent field calibration to translate sensor response to respirable PM concentration. Briefly, 100 sensors underwent a six-point laboratory calibration with dried salt particles. From these experiments, we selected 50 sensors with the most similar calibration slopes (all 50 sensors within ±14% of the average slope of all 100 sensors) for the network (Sousan et al., 2018). Three PM sensors were selected for a field PM calibration procedure with a nephelometer (pDR-1000, Thermo Scientific, Franklin, MA, USA). The three sensors and pDR-1000 were deployed on the same I-beam in the facility for the duration of the study. The mean slope of the PM sensor responses was correlated to the aerosol concentration measured by the pDR-1000 using ordinary least squares linear regression. Five times during the deployment we collected 6–8 h gravimetric respirable dust filter samples at the same location as the pDR-1000 and three collocated sensors. A gravimetric correction factor was derived from each of the filter-based mass concentrations and the pDR-1000, and the average of these five gravimetric correction factors was applied to correct the PM sensor data. The field calibration slope derived from this procedure was then applied to all of the PM sensors in the network for the duration of the study.

For the CO and oxidizing gas sensors, we developed calibration curves in the laboratory using a sample of three sensors of each type (Afshar-Mohajer et al., 2018). Briefly, we exposed the sensors to concentrations of the target gases in a chamber and correlated sensor response with a reference instrument. Although we developed calibration curves for the OX-B421 sensors with both ozone (O₃) and nitrogen dioxide (NO₂) because the sensor responds to both gases without discrimination, we applied the calibration slope for O₃. The calibration curves generated from a sample of three sensors of each type were then applied to all sensors of that type in the monitor network for the duration of this study.

The noise sensor developed for this monitor network used a Microprocessor (Teensy 3.2, open source) with an omnidirectional condenser microphone (CMA-4544PF-W, CUI Inc., Tualatin, OR, USA) (Hallett et al., 2018). Briefly, each of the sensors were calibrated by playing ‘pink noise’ with an acoustic generator (TalkBox, NTi Audio AG, Liechtenstein) and an amplifier (Fender Musical Instruments Corp., Scottsdale, AZ, USA) between 65 and 95 dB in 5 dB increments. Sensor response was compared to the collocated reference sound level metre (XL2, NTi Audio AG, Liechtenstein) for an acceptance criterion of ±2 dB. The results of this calibration procedure were applied for the duration of this study.

Monitor network deployment

The network was installed within an 806 400 square-foot (74 917 m²) area of a more than 2-million square-foot (185 806 m²) manufacturing facility that produces heavy vehicles for construction and forestry. The monitors in our network were deployed in the facility in a spatially optimized pattern to capture maximum spatial variability and reduce monitor redundancy (Berman et al., 2018). Briefly, previous surveys at the facility examined the spatial variability and correlation structure of contaminant concentrations, and an algorithm emphasizing locations to capture temporal variability and high prediction precision was applied to determine the optimal locations for the monitors of our network on the manufacturing floor. From this work, locations were evaluated for a reduced number of monitoring locations that will produce optimal hazard maps for occupational exposure purposes. Forty monitors were deployed at 38 locations on regularly spaced structural I-beams throughout the manufacturing floor in locations closest to those optimal locations as was practicable. Examples of instances where a monitor was not placed at the optimal location included lack of a power outlet for the monitor or I-beam inaccessibility due to obstruction by equipment or construction. An inventory of the manufacturing processes surrounding each monitor was used to group them. The groups and the number of monitors in each group were: machining (5), machining and welding (9), manual welding (11), manual welding and robotic welding (5), staging (2), shot blasting (4), flame cutting (1), shot blasting and laser cutting (3). One central location in the facility was chosen to collocate three monitors, allowing us to evaluate the precision of each type of sensor. At this location we also performed the field calibration routine for the PM sensors.

Data processing

All data analysis was performed with MATLAB R2017a (Natick, MA). We identified and removed database measurements from malfunctioning sensors, identified by ‘flatlined’ or abnormal signals. For each of the sensor types in the network, we imputed all sensor measurements that were below the sensor limit of detection (LOD) with a value of the LOD divided by the square root of 2. The PM sensor LOD was 0.026 mg m⁻³ (Sousan et al., 2018). The LOD for the CO sensor was 10 ppb according to the manufacturer. We estimated the LOD of the OX-B421 sensors in this network was equal to 4 ppb during laboratory evaluations. Briefly, we exposed three OX-B421 sensors to zero air six times for 10 min and calculated the LOD for each of the three i sensors, LOD_OX-B421,i, according to:

$${\displaystyle \begin{array}{l}{\displaystyle \mathrm{L}\mathrm{O}{\mathrm{D}}_{\mathrm{O}\mathrm{X}\text{-}\mathrm{B}421,i}=\frac{3\times {\sigma }_{\mathrm{O}\mathrm{X}\text{-}\mathrm{B}421,i}}{m_{\mathrm{O}\mathrm{X}\text{-}\mathrm{B}421,i}}}\end{array}}$$ (1)

where σ_OX-B421,i is the standard deviation (SD) of the 1-min average responses of the ith OX-B421 sensor in zero air and m_OX-B421,i is the calibration slope of the of the ith OX-B421 sensor to ozone. We averaged the LODs for each of the three sensors to estimate the mean LOD for all sensors in the network. The noise sensor LOD was taken as 65 dBA because the bias above that level in laboratory studies was less than three percent (Hallett et al., 2018). The relatively high LOD of these PM sensors (0.026 mg m⁻³) and the high variability observed during Sunday night periods when PM concentration was lowest, suggested the measurements were not reliable. We therefore took an additional step to correct the PM sensor drift over the 8-month deployment by deriving sensor-specific weekly field calibration intercepts for each sensor on every Sunday night. For the ith week of the study the intercept for the jth sensor was:

$${\displaystyle \begin{array}{l}{\displaystyle \mathrm{I}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{c}\mathrm{e}\mathrm{p}{\mathrm{t}}_{i,j}=\mathrm{m}{\mathrm{V}}_{i,j}-(\mathrm{L}\mathrm{O}{\mathrm{D}}_{\mathrm{P}\mathrm{M}}\times \mathrm{S}\mathrm{l}\mathrm{o}\mathrm{p}{\mathrm{e}}_{\mathrm{P}\mathrm{M}})}\end{array}}$$ (2)

where mV_i,j is the average response in mV, of the jth PM sensor from 00:00 to 02:00 on the ith Sunday, LOD_PM is equal to 0.026 mg m⁻³, and Slope_PM is the field calibration slope used for all PM sensors in the network in mV per mg m^-3 .

Moving 1-h averages of 5-min data were taken and all further calculations were performed on these 1-h averages except where otherwise noted. Data were grouped according to manufacturing processes occurring within a radius of ≤18 m of each monitor. We created violin plots to examine the within- and between-group variability of hazards as well as the distribution of hazard levels within groups. To evaluate the differences in hazard levels between groups of monitors, we implemented a linear mixed effects model. We created daily averages of each hazard for each monitor during typical production periods, nested individual monitors within manufacturing groups, and included the day as a random effect. We adjusted the P-value for multiple comparisons using a Bonferroni correction (alpha level of 0.05; 28 comparisons; P_adjusted = 0.0018). To construct hazard maps, 1-h means from each monitor were plotted and an inverse distance weighting routine was used to interpolate the hazard level at unmeasured locations. These maps were compiled into videos that displayed time and space trends in hazard levels in the study area.

Network precision

Precision of network hazard estimates from the three collocated monitors were examined in two ways. We plotted the difference between each individual monitor and the mean of the three monitors against the mean of the three monitors, which displays the range of hazard estimates at a given concentration or SPL. We also plotted the second-order coefficient of variation (V₂) (Kvålseth, 2017) against the mean of the three monitors for 5-min and 1-h average network measurements. The coefficient of variation (V) is defined as:

$${\displaystyle \begin{array}{l}{\displaystyle V=\frac{\sigma }{¼}}\end{array}}$$ (3)

where σ is the SD and μ is the mean response of collocated monitors. The second-order coefficient of variation was calculated as:

$${\displaystyle \begin{array}{l}{\displaystyle V_2={\left(\frac{V^2}{1+V^2}\right)}^{{}^1╱{}_2}}\end{array}}$$ (4)

V ₂ has bounds from 0 to 1 and approximates the coefficient of variation up to a value of about 0.45 (where V = 0.50). Beyond a value of 0.45, V and V₂ increasingly diverge. We plotted V₂ against the mean to display variation of sensor measurements in a format comparable to coefficient of variation able to accommodate the high levels of variability observed at concentrations near zero.

Network accuracy

We compared hazard measurements from each node of the network to measurements from direct-reading reference instruments in August, December, and March to assess the accuracy of the monitor network. The comparison consisted of collocating measurements with each monitor and reference instruments for 1 min on the first occasion and for 5 min on the second and third occasions. We bypassed the database for this procedure in order to collect 2-s data directly from each monitor via serial connection with a computer. The reference instruments were as follows: respirable PM, personal DataRAM 1500 configured for respirable dust sampling (pDR-1500, Thermo Scientific, Franklin, MA, USA); CO, Q-Trak 7575 (TSI Inc., Shoreview, MN, USA); Personal Ozone Monitor (‘POM’, 2BTechnologies, Boulder, CO, USA); and noise, model XL2 (NTi Audio AG, Liechtenstein). For each monitor, we computed the mean signal from each sensor and converted the output to concentration for the PM, CO, and oxidizing gas sensors using calibration protocols described above (the noise sensor output was dBA). Bias, B, was calculated for each hazard with respect to a reference instrument for each monitor according to:

$${\displaystyle \begin{array}{l}{\displaystyle B=\frac{C_S}{C_T}-1}\end{array}}$$ (5)

where C_S is the mean hazard level measured by the low-cost sensor and C_T is the ‘true concentration’ of the hazard level measured by the reference instrument (NIOSH, 2012). Percent biases were plotted against the reference instrument concentration or SPL to evaluate accuracy across the range of hazard levels observed during the validation campaigns and evaluated against the NIOSH accuracy criterion of bias within ±10% (NIOSH, 2012).

Results

Temporal variability

The multi-hazard monitor network was continuously deployed for 5 months (4 August 2017 to 27 March 2018) and recorded over 2.12 million measurements of PM, CO, oxidizing gases and noise to the database. Over this period of time, the network captured the diurnal and weekly patterns of all hazards (Fig. 1; grey lines represent individual monitors and the black line shows the mean of all monitors). We collected between 166 072 and 179 870 hourly measurements for each hazard, and the percent of measurements that were below the LOD were 3.7%, 3.9%, 0.5%, and 0.2% for PM, CO, oxidizing gases, and noise, respectively. The temporal variability observed for each hazard was consistent with manufacturing activities in the facility, such as daily peaks, decreases during overnight periods and weekend low concentrations. The mean hazards were also correlated with one another (Pearson’s correlation coefficient 0.45 ≤ r ≤ 0.86; see Supplementary Figure S1, available at Annals of Occupational Hygiene online), despite differences in their accumulation, distribution, and dissipation. The mean daily maximum 1-h PM concentrations ± SD recorded across all monitors was 0.62 ± 0.2 mg m⁻³ on typical production days, with individual monitors recording measurements up to 5.9 mg m⁻³. The mean weekend low ± SD was 0.14 ± 0.03 mg m⁻³. The mean daily maximum 1-h CO concentrations ± SD observed across all monitors was 7 ± 2 ppm, with some network nodes occasionally reaching the 13 ppm ceiling of the sensor in this network, and mean concentrations of 1.5 ppm on weekends. The mean daily maximum 1-h oxidizing gas concentrations ± SD observed across all monitors in the network ranged from 155 ± 58 ppb with individual monitors recording measurements of up to 854 ppb O₃ + NO₂ and mean weekend lows ± SD equal to 45 ± 9 ppb. The mean daily maximum 1-h noise SPLs ± SD recorded by all monitors in the network was 82 ± 1 dBA with individual monitors detecting up to 93 dBA and lows on the weekend of 74 ± 1 dBA. The mean daily maximums of PM and CO were normally distributed, but oxidizing gases and noise were not.

Figure 1.

Time series of 1-h average hazard concentrations/intensities measured by the multi-hazard monitor network for (a) PM, (b) CO, (c) O3, and (d) noise. Grey shaded lines are measurements from each individual monitor and black lines display the mean of all monitors.

Intra- and inter-group variability

The variability of hourly PM, CO, oxidizing gases, and noise within and between the groups of work processes for typical production periods, as well as the statistically significant differences in daily hazard means between manufacturing areas from the linear mixed model are displayed in Fig. 2. For PM there was a difference in daily mean concentrations of 0.33 mg m⁻³ between the Machining and Welding (M&W) and Shot Blasting (SB) areas and 0.38 mg m⁻³ between Machining and Welding (M&W) and Shot Blasting & Laser Cutting (SB&LC). There were moderate differences in the variability of PM concentrations within the groups of work processes. For example, the interquartile range (IQR) of PM concentrations in the shot blasting and laser cutting area (SB&LC) was 0.27 mg m⁻³, compared with 0.36 mg m⁻³ in the machining and welding area (M&W). For CO there was a difference in daily mean concentration of 1 ppm between Manual Welding (MW) and Machining and Welding (M&W) areas and 2 ppm between Manual Welding (MW) and Shot Blasting & Laser Cutting (SB&LC) areas. The lowest CO variability was observed in the machining and welding (M&W) area (IQR = 2 ppm) compared with the shot blasting and laser cutting (SB&LC) areas (IQR = 4 ppm). We did not observe statistically significant differences between manufacturing groups for oxidizing gases. The groups of monitors with the greatest oxidizing gas concentration was manual and robotic welding (MW&RW) (133 ppb) and the lowest median oxidizing gas concentration was shot blasting and laser cutting (SB&LC) (90 ppb). With oxidizing gases as well, we observed variability in the distributions of concentrations measured by the network. The shot blasting (SB) area had an IQR equal to 61 ppb, whereas in the shot blasting and laser cutting (SB&LC) area the IQR was equal to 86 ppb. For noise there was a statistically significant difference in daily mean SPLs of 4 dBA between Machining and Welding (M&W) and Staging (S) areas. Noise variability throughout the facility during typical production times was also low, with median SPLs between groups of monitors between 79 and 82 dBA. The variability of noise within groups was similar, with IQRs of SPLs in all manufacturing areas ranging from 2 to 4 dBA.

Figure 2.

Distribution of hazard concentration/intensity by manufacturing processes for (a) PM, (b) CO, (c) O3, and (d) noise. Typical 1-h measurements during production hours from 14 August to 27 March 2018 (time excludes weekends, holidays, and slowdown/shutdown periods). Monitors are grouped by major work processes occurring within a 24 × 37 m area surrounding each monitor. Dashed lines indicate statistically significant differences in daily average hazard levels between manufacturing processes. Manufacturing process abbreviations: machining (M), machining and welding (M&W), manual welding and robotic welding (MW&RW), staging (S), shot blasting (SB), flame cutting (FC), and shot blasting and laser cutting (SB&LC).

Spatiotemporal variability

Examples of hazard maps created for each hazard with data from the network during production periods are displayed in Fig. 3. Videos of each hazard for the period of time between 17 and 18 August 2017 are presented online (see Supplementary Videos S2–S5, available at Annals of Occupational Hygiene online). There was a daily pattern observed for each hazard’s accumulation, peak, and dissipation. For example, PM concentrations originate, spread from and remain highest in areas with machining and welding as primary work processes as seen in the upper centre of the hazard maps. Machining in this facility occurs on a large scale with substantial amount of cutting oil contributing to the PM. For CO, concentrations are highest first, then spread, from areas in the lower left quadrant of the facility where cutting and shot blasting are performed, cutting is a combustion process and may produce CO. Oxidizing gases originate from areas of the facility where many manual welding stations are present (in the centre and upper right quadrant of the hazard maps), mix throughout the facility and dissipate quickly during breaks and at the end of production shifts. Welding arc produces both NO₂ and O₃. Noise contrasts with these spatiotemporally heterogeneous patterns of PM, CO, and oxidizing gases. Noise in the facility increased uniformly throughout the facility from mean overnight lows of 79 dBA to a mean of 83 dBA throughout production times. In contrast to the other hazards under study where specific manufacturing processes contribute to hazard levels, noise is produced everywhere and by nearly all processes in the facility. Noise is a physical hazard that does not disperse from a source the same ways that particulate or gaseous hazards do, leading to the more homogenous SPLs observed in this study. The network has limited ability to capture impact or impulse noise (sudden, brief SPLs exceeding 140 dB) because of the 5-min average SPL recorded to the database; however, examination of 5-min average SPL data (not shown here) does show brief spatially restricted increases in SPL.

Figure 3.

One-hour hazard maps on the morning of 17 August 2017; the day of August validation routine. (a) PM, (b) CO, (c) O3, and (d) noise. For each hazard, the map on the left shows concentrations/intensities before the shift starts and the plot on the right shows concentrations during work operations. Circles represent locations of network nodes (see Supplementary Videos S2–S5, available at Annals of Occupational Hygiene online).

Network precision

Among collocated monitors, we observed 95% of measurements were within 0.21 mg m⁻³ for PM, 0.4 ppm for CO, 9 ppb for oxidizing gases, and 1 dBA for noise (Fig. 4). We observed slightly smaller differences between the three monitors at lower concentrations of CO and oxidizing gases. For noise, the difference in SPL measurements among the three monitors was larger in the higher range of SPLs observed. The second-order coefficient of variation for all hazards is displayed in Fig. 5. The median V₂ of 1-h average measurements for PM, CO, oxidizing gases, and noise of the three collocated monitors was 0.18, 0.02, 0.01, and 0.004, respectively. For all hazards over the duration of the study period, the V₂ did not show a long-term trend and only showed daily and weekly patterns associated with hazard levels. The collocated PM sensors displayed the greatest variability of the four hazards across the range of observed concentrations, and at concentrations of 0.2 and 0.4 mg m⁻³, the V₂ was approximately equal to 0.17 and 0.10, respectively. In general, there was a smaller range of V₂ at a given concentration for 1-h average measurements compared to the 5-min data.

Figure 4.

Precision of 1-h average measurements among collocated monitors for (a) PM, (b) CO, (c) O3, and (d) noise. Each colour represents a different monitor.

Figure 5.

Second-order coefficient of variation (V₂) plotted against mean measurement of three collocated sensors for 5-min (grey) and 1-h (black) averaging time for (a) PM, (b) CO, (c) O3, and (d) noise.

Network accuracy

The bias of the monitors collocated with field reference instruments is shown in Fig. 6, and varied among the hazards under study. The magnitude of the median percent bias between network monitors and field reference instruments were equal to 27%, 11%, 45%, and 1%, for PM, CO, O₃, and noise, respectively. For PM, we observed the magnitude of the percent bias decrease rapidly from a high of 163% with increasing concentrations, with 20% of the measurements meeting the NIOSH bias criteria of percent bias within 10%. For concentrations of CO up to 13 ppm, the magnitude of the bias 92% of collocated measurements were within 25%, and 49% of collocated measurements met the NIOSH bias criterion. For concentrations greater than 13 ppm CO, the observed bias was greater because the concentration was beyond the CO concentration ceiling for these sensors as operated in our network. We observed the magnitude of the bias associated with oxidizing gas concentrations from the OX-B421 sensor and the O₃ reference instrument ranging from 2.1% to 156%. Some of the bias is explained by the fact that the OX-B421 sensor responds to both NO₂ and O₃ with CO₂ as a major interferent (Lewis et al., 2016), compared with the POM reference instrument which is highly specific to O₃. We experienced a malfunction of the POM preventing O₃ bias calculations for the December collocation measurements. In this reduced number of collocated measurements, 9% met the NIOSH bias criterion. The noise sensor outperformed the PM, CO, and oxidizing gas sensors with respect to the magnitude of the bias, which was between 0.13% and 5.70% over the range of SPLs observed in the collocated measurements, and all (100%) measurements met the NIOSH bias criterion.

Figure 6.

Sensor measurement accuracy is shown as %Bias against the concentration/intensity measured by the reference instrument for (a) PM, (b) CO, (c) O3, and (d) noise. Circles from 17 August 2017 with 1-min collocated measurements, squares from 21 and 22 December 2017 with 5-min collocated measurements, and triangles from 23 and 26 March 2018 with 5-min collocated measurements.

Discussion

To our knowledge, this is the first multi-hazard monitor network constructed with low-cost sensors deployed in an industrial setting. This study demonstrates the ability of sensor networks to capture the temporal and spatial patterns of occupational hazards that traditional industrial hygiene approaches with cumulative sampling approaches would not. The hazard maps that were produced with data from the network offer insight into the sources, areas of high concentration, variability in concentrations near different activities, and distribution of hazards. A sensor network could also be used to evaluate if control strategies are effective, offering another advantage over traditional industrial hygiene approaches. Although we observed statistically significant differences between hazard levels and monitors surrounded by different manufacturing processes, we caution against over-interpretation. Certainly the large sample size contributed to the statistical power of our analysis, and we distinguish meaningful differences from statistically significant differences. In practice, an industrial hygienist or occupational health professional might consider the PM and noise differences between manufacturing process areas to be meaningful, whereas the CO differences probably would not be.

In contrast to networks deployed in the ambient environment, this multi-hazard network was deployed in a setting where the distances between monitors were small (less than 41 m with a mean distance to nearest monitor ± SD = 28 ± 8 m), and the concentration of pollutants was high. PM concentrations, for example, in this facility were high enough to foul some sensors and cause signal baseline drift which required correction after relatively short periods of time (Thomas et al., 2018). Another challenge with low-cost sensors that may be exacerbated by high concentration environments is that the variability in sensor measurements may be greater than that of the mean levels of the pollutant under study (Lewis and Edwards, 2016).

A major challenge in mapping occupational hazards is that temporal and spatial variability both contribute to measurement uncertainty (Koehler and Volckens, 2011). As we have demonstrated here, sensor and monitor networks have the potential to overcome this challenge and provide highly temporally and spatially resolved measurements of pollutants—our network recorded PM, CO, oxidizing gas, and noise levels at 5-min time intervals and was spatially optimized to reduce uncertainty of hazard measurements. A major pitfall of hazard mapping is a lack of data completeness, which may lead to incorrect conclusions, underperforming control interventions or wasted resources for surveillance and measurement (Koehler and Volckens, 2011). Such consequences can be ameliorated by using sensor or monitor networks, which address lack of completeness due the nature of a network’s individual nodes being distributed throughout the facility and ability to take measurements simultaneously.

In this study, diurnal and weekly patterns of occupational hazards became apparent after several weeks of deployment, and the full 8 months of data were not needed to establish the repeating patterns of hazards in this facility. In work environments with relatively constant or regular levels of production, a guideline of a 1-month network deployment could be reasonably used to establish spatiotemporal patterns of occupational hazards. We caution against attempting to infer such patterns during times of reduced production, such as holidays or other shutdowns. In our long-term network deployment examples of these periods of time are clearly visible in Fig. 1 including the August manufacturing shutdown when the network was initially deployed, the American Thanksgiving holiday in late November and the Christmas/New Year holiday shutdown in late December and early January. Surprisingly, we did not observe a strong seasonal influence on hazard levels, which might reasonably be expected due to changes in heating, ventilation, and air conditioning (HVAC) practices in the facility. An advantage of the long-term deployment of a sensor or monitor network is that this kind of long-term temporospatial variability can be explicitly characterized in a way that intermittent measurements or measurements at a limited number of locations cannot. Another advantage offered by our network was the ability to observe the spatial variation of hazards and associate their concentrations with specific manufacturing processes.

The use of low-cost sensors in wireless networks poses many challenges including the need for thorough laboratory and field calibration, sensor baseline drift over time, and overall lower data quality (Snyder et al., 2013; Piedrahita et al., 2014; Xiong and Compton, 2014; Masson et al., 2015; Lewis and Edwards, 2016; Lewis et al., 2016). Another challenge, for gas sensors in particular, is sensor specificity to the target gas, which produces erroneous response from interfering gases (Masson et al., 2015; Spinelle et al., 2015). In this network, the oxidizing gas sensor signal is a summation of response to both O₃ and NO₂, and is unable to discriminate between the two gases (Hossain et al., 2016; Afshar-Mohajer et al., 2018). The oxidizing gas sensor’s non-specific response makes accurate estimation of ozone quite challenging and a comparison to the PEL difficult at best and misleading at worst. For example, a similar response from the OX-B421 sensor to NO₂ at the mean to maximum levels of response observed in this study in the absence of O₃ would be associated with approximately 0.1–0.5 ppm of NO₂, well below the PEL (ceiling) of 5 ppm NO₂. Future multi-hazard monitor networks may attempt to improve O₃ concentration estimates by employing pared electrochemical sensors, one oxidizing gas sensor (O₃ + NO₂), like was used in our network, and one sensor specific to NO₂ (Hossain et al., 2016).

Sensor calibration in sensor networks with a large number of sensor nodes is a key consideration in their deployment. Users have three main calibration options: (i) users can apply the manufacturer’s calibration constants for slope and intercept, (ii) create a calibration curve specific to each sensor in the field or in the laboratory, or (iii) develop a calibration curve that can be generalized to all sensors in a given lot of sensors in the field or laboratory. The advantage of applying a common calibration curve to all sensors in a network includes a simplification of data processing and translation of sensor signal to concentration and the avoidance of calibrating each sensor in the sensor network. On the other hand, using a common calibration curve based on a sample of sensors introduces a source of measurement error in a sensor network because of variability in the response of sensors of the same type. We suggest calibrating a sample of sensors set up in their intended configuration and evaluating if the variability in the calibration slope of the sample of sensors exceeds the tolerance of acceptable measurement error for a given application. This strategy does however require some prior knowledge about the concentration of target gas in the environment of interest and is the subject of future work.

Unfortunately, none of these three sensor calibration strategies solve or take into account that the calibration slope of some types of sensors change over time (Afshar-Mohajer et al., 2018), calibration relationships may only hold for specific locations for a limited period of time (Lewis and Edwards, 2016), or calibration may differ substantially in the laboratory versus the field (Piedrahita et al., 2014). A limitation of this study is that the same calibration slopes were used throughout this study for the PM, CO, and OX sensors, which likely introduced increasing measurement error over time. Future work should consider how in-field assessment can be used to update the calibration of low-cost sensors, for example with the collocation of a higher-quality field reference instrument or the use of calibration gases.

Concentrations of indoor contaminants are highly variable, autocorrelated in time and space, and related to occupant activities, which complicates statistical procedures and likely leads to an underestimation of variance (Francis et al., 1989; Symanski and Rappaport, 1994; Høst et al., 1995; Luoma and Batterman, 2000; Kolovos et al., 2010). Although a variety of statistical techniques have been applied to address these issues of non-independence and autocorrelation, in this preliminary data analysis we have interpolated hazard concentrations between measured locations, using inverse-distance weighting, a technique that ignores spatial and temporal autocorrelation in the mapped measurements (Koehler and Peters, 2013). Future work will examine the correlation structure of the hazards measured with this network and characterize the statistical variability in mapped hazard levels with strategies such as Kriging.

Conclusions

Here we demonstrated the ability of a spatially dense (maximum distance to the nearest monitor equal to 41 m) sensor network to collect information over 8 months on multiple occupational hazards at a time interval of 5 min. Examination of network data provided insight into the daily, weekly, and seasonal patterns and the spatial distribution of hazards in the facility including hotspot identification that would not be possible with traditional industrial hygiene approaches. It also allowed us to examine the manufacturing processes associated with higher levels of the various hazards in the network. Despite these successes, serious challenges with sensor accuracy, precision, stability over time, and cross-sensitivity to non-target species persists. In campaigns to verify the accuracy of the network, we observed a range of bias with respect to high-quality direct reading instruments depending on the hazard and the concentration/level, with median biases ranging from 1% for noise to 41% for PM. Within a set of three collocated monitors in the network, we observed a range in precision by hazard and absolute differences between monitors that tended to be greater at higher hazard levels and relative differences between monitors that were higher at low hazard concentrations. These lessons learned in this study, as well as the account of our experience are generalizable to others who wish deploy sensor networks in occupational environments. Although different workplaces will have different hazards of interest, this study offers an example for future work. In addition, the approaches we present in this study can be applied to an expanding catalogue of hazards for which there is a growing number of commercially available sensors on the market. Future work will investigate the feasibility of using a sensor network to quantitatively estimate personal exposure in the occupational environment.

Funding

Funding for this project was provided by NIOSH grant R01OH010533. C.Z. was supported by the Johns Hopkins University Education and Research Center for Occupational Safety and Health (ERC), which is funded by NIOSH under grant number 5 T42 OH 008428.

Conflict of Interest

The authors declare no conflict of interest relating to the material presented in this article. Its contents, including any opinions and/or conclusions expressed, are solely those of the authors.