Design and Testing of a Low-Cost Sensor and Sampling Platform for Indoor Air Quality

Abstract

Americans spend most of their time indoors at home, but comprehensive characterization of in-home air pollution is limited by the cost and size of reference-quality monitors. We assembled small “Home Health Boxes” (HHBs) to measure indoor PM_2.5, PM₁₀, CO₂, CO, NO₂, and O₃ concentrations using filter samplers and low-cost sensors. Nine HHBs were collocated with reference monitors in the kitchen of an occupied home in Fort Collins, Colorado, USA for 168 h while wildfire smoke impacted local air quality. When HHB data were interpreted using gas sensor manufacturers’ calibrations, HHBs and reference monitors (a) categorized the level of each gaseous pollutant similarly (as either low, elevated, or high relative to air quality standards) and (b) both indicated that gas cooking burners were the dominant source of CO and NO₂ pollution; however, HHB and reference O₃ data were not correlated. When HHB gas sensor data were interpreted using linear mixed calibration models derived via collocation with reference monitors, root-mean-square error decreased for CO₂ (from 408 to 58 ppm), CO (645 to 572 ppb), NO₂ (22 to 14 ppb), and O₃ (21 to 7 ppb); additionally, correlation between HHB and reference O₃ data improved (Pearson’s r increased from 0.02 to 0.75). Mean 168-h PM_2.5 and PM₁₀ concentrations derived from nine filter samples were 19.4 μg m⁻³ (6.1% relative standard deviation [RSD]) and 40.1 μg m⁻³ (7.6% RSD). The 168-h PM_2.5 concentration was overestimated by PMS5003 sensors (median sensor/filter ratio = 1.7) and underestimated slightly by SPS30 sensors (median sensor/filter ratio = 0.91).

Graphical Abstract

1. INTRODUCTION

Exposure to air pollution is a leading risk factor for human morbidity and mortality [1]. To protect human health, government agencies, such as the United States Environmental Protection Agency (US EPA), have promulgated standards for outdoor concentrations of air pollutants including particulate matter (PM), carbon monoxide (CO), nitrogen dioxide (NO₂), and ozone (O₃) [2]; however, Americans typically spend 60% to 85% of their time at home and, as a result, more than half of personal exposure to air pollutants such as fine particulate matter (PM_2.5) typically occurs in the indoor home environment [3-7]. Outdoor air pollution (e.g., from industrial sources, motor vehicles, and wildfires) can enter the home via open windows, mechanical ventilation systems, and infiltration of the building envelope [8]. Additionally, indoor sources can produce airborne pollutant concentrations that exceed outdoor levels [8-11]. For example, use of natural gas-fueled cooking burners increases indoor CO₂, CO, NO_x, and particle number concentrations [9-11]. Singer et al. found that use of gas-fueled burners and ovens during simulated cooking activities in homes with the windows closed, forced-air ventilation turned off, and the range exhaust hood turned off often resulted in 1-h average kitchen NO₂ concentrations above the US EPA outdoor 1-h standard of 100 ppb [10]. Even when electric or induction burners are used, PM_2.5 emitted from cooking food can lead to peak indoor concentrations > 100 μg m⁻³ [12,13].

Time-resolved data on home air quality can inform occupants, along with health and air quality professionals, about trends in and sources of air pollutants in a specific home and might help them take direct action, seek resources, and/or advocate for policies to mitigate those pollutants; however, efforts to collect such data are limited by the costs of instrumentation and sample analysis. Low-cost sensors can reduce the hardware costs associated with acquiring time-resolved pollutant concentration data, but low-cost sensor data must be interpreted with care. For example, the inaccuracies associated with sensors that convert particle light scattering signals to PM mass concentrations are well-documented [14-20]. The relationship between scattered light intensity and aerosol mass concentration depends on the shape, refractive index, size distribution, and density of the particles [21]. Without a mechanism for drying incoming particles, low-cost sensors can overestimate dry PM mass concentrations when hygroscopic particles take up water in high humidity environments [18,22]. In addition, many low-cost light scattering sensors fail to respond to increases in coarse (2.5 to 10 μm) particle mass—most likely because of particle losses in the flow path between the inlet and the light source/detector as well as decreases in the amount of light scattered over the angular detection range per unit mass [16,20,23,24].

Several recent publications have documented the challenges associated with converting raw outputs from low-cost NDIR CO₂ sensors—as well as electrochemical CO, NO₂, and O₃ sensors—into accurate pollutant concentrations. Most studies describe outdoor monitoring applications in which temperature- and humidity-sensitive gas sensors were subjected to variable environmental conditions coupled with low pollutant concentrations [25-32]. In many of these studies, low-cost sensors were collocated with reference (i.e., research- or regulatory-grade) monitors for several months and the collocation data were then used to train linear and/or machine learning calibration models. Though effective, these calibration approaches present barriers to users who lack access to reference monitors and increase the effective “cost” of the sensors.

Few publications have described indoor monitoring with low-cost gas sensors [33,34]. We are not aware of any studies in which (a) low-cost CO₂, CO, NO₂, and O₃ sensors were collocated with a comprehensive suite of reference instruments in an indoor environment or (b) data on the efficacy of sensor manufacturer-supplied calibrations were reported. Low-cost gas sensors might provide more reliable data indoors, where temperature and relative humidity (RH) vary less than outdoors and where indoor sources can elevate CO₂, CO, and NO₂ concentrations above outdoor levels.

In the present study, we developed a prototype platform—called the “Home Health Box” (HHB)— that used time-integrated pollutant samplers and low-cost sensors to measure PM_2.5, PM₁₀, CO₂, CO, NO₂, O₃, and volatile organic compound (VOC) concentrations. We deployed nine HHBs alongside reference monitors in the kitchen of an occupied home in Fort Collins, CO, USA for one week to investigate: (1) how CO₂, CO, NO₂, O₃, and PM_2.5 concentrations in the kitchen varied with outdoor pollution levels and in-home activities; (2) whether CO₂, CO, NO₂, and O₃ concentrations calculated using the low-cost sensor manufacturers’ calibration data agreed with concentrations reported by reference monitors; (3) whether linear calibration models developed through collocation with reference monitors improved the accuracy of CO₂, CO, NO₂, and O₃ concentrations derived from low-cost sensor data; (4) the accuracy and precision of PM concentrations measured using filter samplers and low-cost sensors in collocated HHBs; (5) whether reference monitor and low-cost sensor data led to the same conclusions about whether pollutant concentrations were low, elevated, or high relative to various air quality standards; as well as (6) whether reference monitor and low-cost sensor data led to the same conclusions regarding which in-home activities contributed disproportionately to the time-integrated concentration of each pollutant.

2. MATERIAL AND METHODS

2.1. Home Health Box

The HHB (Figure 1) was designed to sample PM_2.5, PM₁₀, and VOCs while measuring PM, CO₂, CO, NO₂, and O₃ concentrations continuously. Inside each HHB, two modified ultrasonic personal aerosol samplers (UPAS, Access Sensor Technologies, Fort Collins, CO, USA) collected PM_2.5 and PM₁₀ onto 37-mm PTFE filters (PT37P-PF03, MTL, Minneapolis, MN, USA) at 2 L min⁻¹ [35,36]. A third modified UPAS sampled VOCs onto a thermal desorption tube at 0.003 L min⁻¹. Two low-cost sensors (PMS5003, Plantower, Beijing, China and SPS30, Sensirion, Stäfa, Switzerland) employed light-scattering techniques to measure real-time variations in PM concentrations. Concentrations of CO₂ were measured using a low-cost NDIR sensor (SCD30, Sensirion, Stäfa, Switzerland). Concentrations of CO, NO₂, and O₃ were measured using electrochemical sensors (CO-B4, NO2-B43F, and OX-B431 Alphasense, Great Notley, Essex, United Kingdom) mounted to Individual Sensor Boards (000-0ISB-00 and 000-0ISB-02, Alphasense). The CO₂, CO, NO₂, and O₃ sensors were isolated from the rest of the internal components by a 3D-printed housing (Figure S1). This housing had one inlet and one outlet on the outside of the main HHB enclosure. The NO₂ and O₃ sensors were installed immediately downstream of the inlet to minimize losses of these species (but we did not characterize losses). Air flowed through the gas sensor housing continuously with the aid of a small fan installed on the outlet (UF3H3-710, Sunon, Kaohsiung City, Taiwan). Temperature and RH inside the gas sensor housing were measured by a sensor included on the SCD30 circuit board.

Figure 1.

Left: The Home Health Box enclosure had exterior dimensions of 220 × 170 × 130 mm (excluding the feet, sample inlets, exhaust vents, hinge, and latch) and weighed approximately 2 kg. Right: A top view of the Home Health Box with the cover removed to show internal components.

The HHB enclosure (YH-080604, Polycase, Avon, OH, USA) had exterior dimensions of approximately 220 × 170 × 130 mm. The HHB was designed to plug into a 120-V wall outlet but was equipped with a backup battery (V25, Voltaic Systems, Brooklyn, NY, USA) to prevent sample interruption during a short-term (< 6 h) loss of line power. The HHB weighed ~2 kg. Firmware running on an internal microcontroller (STM32L475VGT6, STMicroelectronics, Geneva, Switzerland) controlled sample flow rates, the sample duration, and data logging (Table S1).

2.2. In-Home Experiment

Approval for in-home sampling with the Home Health Boxes was obtained from the Institutional Review Board at Colorado State University (protocol number 19-9443H).

2.2.1. Experimental Setup

The experiment took place in the kitchen of a 235-m² detached house in Fort Collins, Colorado, USA. The house—which was built in 1991—featured a ground floor (where the kitchen was located), a partially-finished basement, an attached two-car garage, and central air conditioning.

The 4.3 × 5.1-m kitchen was enclosed on three sides (north, west, and south) by two exterior walls and one interior wall (Figure S2). The ceiling height ranged from 2.5 m (at the west exterior wall) to 3.6 m (on the east side where the kitchen was connected to the main living area). The wall on the fourth (east) side of the kitchen extended from the floor to a height of 2.2 m (i.e., there was a >1 m gap between the top of this wall and the ceiling). A 1.6-m-wide by 2.1-m-tall opening (with no door) in the east wall served as a walkway between the kitchen and the living area.

Unvented natural gas cooking burners were present in the stove installed along the south wall (MGR7662WQ, Maytag, Benton Harbor, MI, USA). The left front, right front, left rear, and right rear cooktop burner firing rates were specified as 13.5, 12.5, 9.5, and 5.0 kbtu hr⁻¹, respectively. The oven and broiler firing rates were specified as 16.5 and 10.0 kbtu hr⁻¹, respectively. The hood installed above the cooktop (233.5338910, Kenmore, Hoffman Estates, IL, USA) exhausted back into the kitchen, instead of to the outdoors, and was not typically used by home occupants.

A portable air cleaner (AP-1512HH, Coway, Seoul, Korea) was also present in the kitchen. This air cleaner, which was rated for a 33.5 m² room, passed air through a carbon “pre-filter”, a HEPA filter, and a bipolar ionizer. The carbon and HEPA filters were replaced just before the experiment started. A button on the unit allowed the ionizer to be switched on and off separately.

We deployed nine Home Health Boxes in the kitchen to test their ability to measure a wide range of PM, CO₂, CO, NO₂, and O₃ concentrations resulting from indoor sources (e.g., emissions from natural gas cooking burners and cooking food) as well as infiltration of outdoor air pollution. Each HHB wrote timestamped data to a microSD card at 30-s intervals.

Reference measurements of indoor CO₂, CO, NO, NO_x, O₃, and PM_2.5 levels were logged at 60-s intervals using an LI-820 CO₂ Gas Analyzer (LI-COR Biosciences, Lincoln, NE, USA), two QTrak Indoor Air Quality Monitors (7575-X with model 982 probe, TSI Incorporated, Shoreview, MN, USA), a Trace Level Chemiluminescence NO-NO₂-NO_x Analyzer (Model 42C, Thermo Environmental Instruments, Franklin, MA, USA), a UV Photometric O₃ Analyzer (Model 49C, Thermo Environmental Instruments, Franklin, MA, USA), and a Tapered Element Oscillating Microbalance (TEOM 1405, Fisher Scientific, Waltham, MA, USA). The TEOM sample flow rate was set to 4 L min⁻¹ and a GK2.05 (KTL) cyclone (2.5 μm cutpoint at 4 L min⁻¹; Mesa Laboratories, Butler, NJ, USA) was installed on the sample inlet (i.e., the 16.7 L min⁻¹ inlet typically used for regulatory monitoring was not used). See Supporting Information (SI) Section S1 for additional information on how these instruments were calibrated and operated.

We aimed to place the indoor monitors as close together as possible given the constraints imposed by the kitchen layout and the fact that the home was occupied during the experiment. The nine HHBs were stacked three high and arranged into three columns spaced 0.91 m apart on a shelf 1.3 m above the cooktop (Figure 2). The center column was aligned with the cooktop. The LI-820 and the two QTraks were set on top of a microwave 1.1 m to the right of the cooktop (from the perspective of a person standing in front of the oven). The 42C NO-NO₂-NO_x Analyzer and 49C O₃ Analyzer were installed along the north wall. A PTFE sample line leading to these analyzers was secured 1.1 m above and 3.2 m away from the cooktop (Figure S2). The TEOM was placed on the same shelf as the HHBs, 1.9 m to the left of the cooktop.

Figure 2.

Placement of the Home Health Boxes, LI-820 CO₂ Gas Analyzer, QTrak Indoor Air Quality Monitors (used to measure CO), and TEOM 1405 (used to measure PM_2.5) relative to the natural gas cooking burners along the south interior kitchen wall. The >1-m gap between the top of the wall and the ceiling on the east side of the kitchen is visible around the TEOM. The 42C NO-NO₂-NO_x Analyzer and 49C O₃ Analyzer were installed in the same room but are not visible from this perspective. Occupants’ personal effects have been blurred in the photo.

Outdoor particulate matter concentrations and meteorological conditions were recorded at 30-s intervals using a second-generation Aerosol Sensor Plus Environmental Node (ASPEN) Box [17] installed 1.5 m outside the north kitchen exterior wall (Figure S2). This ASPEN Box included a PM_2.5 filter sampler, a PM₁₀ filter sampler, a PMS5003 sensor, an SPS30 sensor, as well as a temperature and RH sensor (SHT30, Sensirion) installed in a radiation shield.

City-level, 10-minute average temperature, RH, and solar radiation data were obtained from the Fort Collins Weather Station located 4.0 km from the home on the Colorado State University main campus [37]. City-level, 1-h ambient CO and O₃ concentrations were obtained from monitors (48i-TLE, Thermo Scientific, Franklin, MA, USA and 400E, Teledyne API, San Diego, CA, USA, respectively) operated by the Colorado Department of Public Health and Environment (CDPHE) at EPA AQS site 08-069-1004 (4.3 km from the home). City-level, 1-h ambient PM_2.5 and PM₁₀ concentrations were obtained from a monitor (EDM180, GRIMM Aerosol Technik, Ainring, Germany) operated by CDPHE at AQS site 08-069-0009 (3.7 km from the home).

2.2.2. Activities

The 168-h experiment spanned October 8 to 15, 2020 UTC. During the experiment, outdoor PM concentrations were elevated due to the Cameron Peak fire burning west of Fort Collins and the Mullen fire burning 130 km northwest of Fort Collins (Figure S3) [38,39]. The Cameron Peak fire grew from 131,000 acres on October 7^th to 164,000 acres on October 14^th [38,40].

The house was occupied by three unrelated adults and two dogs. All adults were employed outside the home, but one worked almost exclusively from home due to the COVID-19 pandemic. During the experiment, all monitors measured air pollution from normal occupant activities as well as scripted activities that were designed to purposefully increase indoor NO₂ or O₃ concentrations.

Occupants kept a log of normal (i.e., non-scripted) activities that might have affected indoor pollutant concentrations (e.g., cooking, opening windows). After the experiment, this log was reviewed alongside time-series reference monitor data to see if occupants could identify any pollution-generating or pollution-depleting events that were missing from the original log.

To modulate the NO₂ concentration in the kitchen, an occupant completed six scripted activities with natural gas cooking burners. The occupant boiled 3 L of water in a pot on the left front cooktop burner on five separate occasions (Table 1): (1) while all windows and doors were closed, the air conditioner was off, and the air cleaner was off; (2, 3) while the air cleaner HEPA filter was on; (4) while two windows and a sliding glass door were open; and (5) while the air conditioner was on. On October 14^th, the occupant operated the broiler for approximately 8 h with the oven door closed and with no food inside the oven. While the broiler was on, all windows and doors to the outside remained closed, the air conditioner remained off, and the air cleaner remained off. This last activity was designed to capture the high indoor NO₂ concentration that could result from use of a gas burner during long-duration cooking (e.g., in preparation for a holiday meal) [41,42].

Table 1.

A summary of the scripted activities that an occupant completed to purposefully increase NO₂ and O₃ concentrations in the kitchen. All times are the local time in Fort Collins, CO, USA (UTC-06:00 on the dates listed). The “HEPA filter” and “Ionizer” columns represent two separate functions of the indoor air cleaner.

Target pollutant	Description	Start time	End time	Door & window	AC	HEPA filter	Ionizer
NO2	Boiled 3 L of water	08 Oct. 11:25	08 Oct. 13:05	Closed	Off	Off	Off
NO2	Boiled 3 L of water	12 Oct. 09:40	12 Oct. 10:20	Closed	Off	On	Off
NO2	Boiled 3 L of water	12 Oct. 10:45	12 Oct. 11:35	Closed	Off	On	Off
NO2	Boiled 3 L of water	12 Oct. 13:30	12 Oct. 15:55	Open	Off	Off	Off
NO2	Boiled 3 L of water	13 Oct. 09:50	13 Oct. 13:05	Closed	On	Off	Off
NO2	Turned on broiler	14 Oct. 10:30	14 Oct. 16:50	Closed	Off	Off	Off
O3	Opened door and window	08 Oct. 14:30	08 Oct. 16:00	Open	Off	Off	Off
O3	Opened door and window	11 Oct. 09:00	11 Oct. 12:00	Open	Off	Off	Off
O3	Turned on ionizer	10 Oct. 13:30	11 Oct. 10:00	Closed	Off	On	On

To modulate the O₃ concentration in the kitchen, an occupant completed two scripted activities in which two windows and a sliding glass door were opened while all cooking burners remained off. The air conditioner and air cleaner were both off on these occasions. To investigate whether the bipolar ionizer increased the O₃ concentration in the kitchen, we scripted a separate activity in which an occupant turned on the HEPA filter and ionizer for 21.5 hours while the doors and windows remained closed and the air conditioner remained off.

2.2.3. Sample Analyses

Ten PM_2.5 filter samples and ten PM₁₀ filter samples (nine HHB + one ASPEN) were collected during the experiment. The 37-mm PTFE filters (PT37P-PF03, Measurement Technology Laboratories, Minneapolis, MN, USA) used to collect these samples were pre- and post-weighed on a balance with 1-μg resolution (XS3DU, Mettler-Toledo, Columbus, OH, USA) [43].

The VOC samples collected using the HHBs were not analyzed and, thus, data on indoor VOC levels are not presented herein.

2.2.4. Data Analyses

2.2.4.1. Reference Monitor Data

The two CO concentrations measured using the two QTrak monitors during each minute were averaged. This average concentration was used as the reference CO concentration. Additionally, at each 60-s interval, the indoor NO₂ concentration was calculated as the difference between the NO_x and NO concentrations reported by the 42C NO-NO₂-NO_x Analyzer.

2.2.4.2. Gas Sensor Calibrations

First, CO, NO₂, and O₃ concentrations were estimated using algorithms and sensor-specific calibration data provided by Alphasense. During this process, the “corrected working electrode voltage” (WE_c; mV) had to be calculated for each sensor serial number i at each 30-s log interval j. Alphasense provided four possible algorithms for calculating WE_c,ij [44]:

$$W{E}_{c,ij}=\left(W{E}_{u,ij}-W{E}_{e,i}\right)-n_{T,ij}\left(A{E}_{u,ij}-A{E}_{e,i}\right)$$ (1)

$$W{E}_{c,ij}=\left(W{E}_{u,ij}-W{E}_{e,i}\right)-k_{T,ij}\frac{W{E}_{0,i}}{A{E}_{0,i}}\left(A{E}_{u,ij}-A{E}_{e,i}\right)$$ (2)

$$W{E}_{c,ij}=\left(W{E}_{u,ij}-W{E}_{e,i}\right)-\left(W{E}_{0,i}-A{E}_{0,i}\right)-k_{T,ij}^{\prime }\left(A{E}_{u,ij}-A{E}_{e,i}\right)$$ (3)

$$W{E}_{c,ij}=\left(W{E}_{u,ij}-W{E}_{e,i}\right)-W{E}_{0,i}-k_{T,ij}^{″}$$ (4)

The pollutant concentration measured by sensor serial number i at time j (c_ij; ppb) was then calculated from the corrected working electrode voltage using Equation 5 [44]:

$$c_{ij}=\frac{W{E}_{c,ij}}{r_{T,ij}{s}_i}$$ (5)

In Equations 1-5, WE_u,ij and AE_u,ij were the uncorrected working and auxiliary electrode voltages recorded by sensor serial number i at time j (mV), WE_e,i and AE_e,i were the working and auxiliary electrode electronic offsets for individual sensor board i (mV); WE_0,i and AE_0,i were the working and auxiliary electrode zero voltages for sensor i (mV); n_T,ij, k_T,ij, k′_T,ij and k″_T,ij were temperature-dependent correction factors; r_T,ij was the relative sensitivity as a function of temperature; and s_i was the sensitivity of sensor serial number i to the relevant pollutant (mV ppb⁻¹). WE_e, AE_e, WE₀, AE₀, and s were calibration constants provided by Alphasense for each sensor serial number i. The correction factors n_T, k_T, k′_T, k″_T, and r_T were provided by Alphasense, in tabular form (n_T, k_T, k′_T, k″_T) or graphical form (r_T), as a function of sensor part number (CO-B4, NO2-B43F, or OX-B431) and temperature [44-47]. These correction factors were evaluated at the temperature measured inside gas sensor housing i at time j (T_ij; °C) by interpolating between values in the table or graph.

Because the OX-B431 sensor responds to both O₃ and NO₂, it must be used in conjunction with the NO2-B43F sensor to determine O₃ concentrations. To calculate O₃ concentrations, all instances of (WE_u,ij – WE_e,i) in Equations 1-4 were replaced with (WE_u,ij – WE_e,i – WE_c,NO2,ij), where WE_c,NO2,ij (mV) was the product of the sensitivity of the OX-B431 sensor to NO₂ (s_NO2,i; mV ppb⁻¹) and the NO₂ concentration calculated from the NO2-B43F data at time j using Equations 4 and 5 (c_NO2,ij) [44]:

$$W{E}_{c,NO2,ij}=c_{NO2,ij}{s}_{NO2,i}$$ (6)

Each gas concentration (CO, NO₂, O₃) was calculated using each of the four algorithms at each 30-s interval in the 168-h experiment. If the concentration (c_ij) calculated using Equation 5 was below 0, the calculated value was replaced with zero.

Second, we fit our own linear mixed models to predict the 1-h average gas concentration measured at time j (c_j; ppb) by the reference monitor(s) for the pollutant of interest (c_CO,j for CO-B4 sensors, c_NO2,j for NO2-B43F sensors, and c_NO2,j + c_O3,j for OX-B431 sensors) from 1-h average working electrode voltages, temperatures, and RH values:

$$c_j=\alpha +a_i+(\beta +b_i)\left(W{E}_{u,ij}-W{E}_{e,i}\right)+{ϵ}_{ij}$$ (7)

$$c_j=\alpha +a_i+(\beta +b_i)\left(W{E}_{u,ij}-W{E}_{e,i}\right)+{\gamma }_T{T}_{ij}+{ϵ}_{ij}$$ (8)

$$c_j=\alpha +a_i+(\beta +b_i)\left(W{E}_{u,ij}-W{E}_{e,i}\right)+{\gamma }_{RH}R{H}_{ij}+{ϵ}_{ij}$$ (9)

where ϵ_ij was the random error. These models included the working electrode voltage as a predictor, and not the auxiliary electrode voltage, because (a) (WE_u,ij – WE_e,i) was strongly correlated with the reference gas concentration for all three sensors and (b) Equation 4, which did not use the auxiliary electrode voltage, was the best-performing Alphasense algorithm for the CO-B4 and NO2-B43F sensors (see Section 3.2.1). Coefficient α was a fixed intercept, β was a fixed slope, a_i was a random intercept for sensor serial number i, and b_i was a random slope for sensor serial number i. The random intercept and slope were included because (a) Alphasense provided a unique sensitivity (si in Equation 5; mV ppb⁻¹) for each serial number, thereby implying that each individual sensor had a unique response to the pollutant of interest, and (b) model results indicated that there was random variation in the intercept and slopes between serial numbers. T_ij and RH_ij were the temperature (°C) and relative humidity (%), respectively, inside gas sensor housing i at time j. The coefficients γ_T and γ_RH remained constant across all sensors with a given part number (CO-B4, NO2-B43F, or OX-B431). Temperature and RH were included as predictors in Equations 8 and 9, respectively, because documents from Alphasense [44-48], and prior studies [25,28-31], indicated that temperature and RH influenced sensor performance.

We fit the linear mixed model shown in Equation 10 to relate 1-h average CO₂ concentrations reported by the LI-820 (c_{CO2,LI–820,j}; ppm) to 1-h average CO₂ concentrations reported by the low-cost NDIR sensors (C_CO2,SCD30,ij; ppm):

$$c_{CO2,SCD30,ij}=\alpha +a_i+\gamma c_{CO2,LI-820,j}+{ϵ}_{ij}$$ (10)

where α was a fixed intercept, a_i was a random intercept for sensor serial number i, γ was a fixed slope, and ϵ_ij was the random error. Sensirion’s documentation for the SCD30 did not suggest that a random slope would be required for each sensor serial number. A model that also included a random slope for each serial number was considered, but the between-serial number variance associated with the random slope was zero—indicating that it was not needed.

The models shown in Equations 7-10 were trained and tested using seven-fold cross validation. Each fold consisted of a continuous 24-h period spanning midnight to midnight UTC. Models fit to the other 144 h of data were used to predict 1-h average concentrations in a fold. This process was repeated seven times to obtain predictions for the full 168-h duration of the experiment. The models shown in Equations 7-9 were used to predict 1-h average CO, NO₂, and NO₂ + O₃ concentrations. The predicted O₃ concentration was calculated by subtracting the predicted NO₂ concentration from the predicted NO₂ + O₃ concentration. The model shown in Equation 10 was used to predict each 1-h average CO₂ concentration by subtracting the intercepts from the sensor-reported CO₂ concentration and dividing the difference by the slope: (c_CO2,SCD30,ij – α – a_i)/γ.

We compared 1-h average concentrations (a) obtained directly from the SCD30 CO₂ sensors in the HHBs, (b) calculated from HHB data using the Alphasense-recommended algorithms shown in Equations 1-4, and (c) predicted from HHB data using the linear mixed models shown in Equation 7-10 to 1-h average concentrations measured by the reference monitors to determine the root-mean-square error (RMSE), mean absolute error (MAE), mean bias error (MBE), normalized mean absolute error (NMAE), Pearson correlation coefficient, precision, and bias for each sensor serial number i (SI Section S1.2.5) [27-29,31,49,50].

Electrochemical gas sensor data from one HHB were excluded from the model fitting process and all results because those data consisted of unusually high working and auxiliary electrode voltages. Working and auxiliary voltages read from the excluded NO2-B43F sensor were always > 2 V, whereas working and auxiliary voltages read from the other eight NO2-B43F sensors were always < 0.4 V. In addition, working electrode voltages read from the excluded CO-B4 sensor were several hundred mV higher than those read from the other eight CO-B4 sensors.

2.2.4.3. PM Sensor Calibrations

For each PMS5003 or SPS30 sensor i, the corrected PM_2.5 concentration measured at 30-s interval j (c_PM2.5,ij; μg ^m−3) was calculated as:

$$c_{PM2.5,ij}=\frac{{\overline{c}}_{PM2.5,filter,i}}{{\overline{c}}_{PM2.5,sensor,i}}{c}_{PM2.5,sensor,ij}$$ (11)

where ${\overline{c}}_{PM2.5,filter,i}$ was the 168-h average PM_2.5 concentration derived from the filter sample collected using HHB i (μg m⁻³), ${\overline{c}}_{PM2.5,sensor,i}$ was the time-averaged PM_2.5 concentration reported by sensor i over the duration of the filter sample (μg m⁻³), and c_{PM2.5,sensor,ij} was the PM_2.5 concentration reported by sensor i at time j (μg m⁻³). The ratio (${\overline{c}}_{PM2.5,filter,i}∕{\overline{c}}_{PM2.5,sensor,i}$) was the gravimetric correction factor for sensor i. For PMS5003 sensors, “PM_2.5 CF=1” values were used as c_PM2.5,sensor.

We compared 1-h average values of (a) c_{PM2.5,sensor,ij} and (b) c_PM2.5,ij obtained from HHB data to 1-h average PM_2.5 concentrations measured using the TEOM during the first 48 h of the experiment to determine the RMSE, MAE, MBE, NMAE, and Pearson correlation coefficient for each sensor serial number i. This analysis only spanned the first 48 h of the experiment because the TEOM stopped operating reliably after that time.

We also compared 24-h average values of c_PM2.5,ij calculated from HHB PMS5003 sensor data using (a) Equation 11 and (b) a United States-wide correction recently developed for outdoor PurpleAir monitors by Barkjohn et al. [51] to 24-h average PM_2.5 concentrations measured by the TEOM during the first 48 h of the experiment (Equation S9). For each of these two correction approaches, the RMSE, MAE, MBE, and NMAE were calculated from the eighteen 24-h PM_2.5 concentrations obtained from the nine HHBs over this 48-h period.

2.2.4.4. Classification of Pollutant Concentrations

Prior to the experiment, we developed a tiered system to classify the level of each pollutant (CO₂, CO, NO₂, O₃, PM_2.5, and PM₁₀) as low, elevated, or high based on guidelines from the World Health Organization (WHO); US EPA; American Society of Heating, Refrigerating, and Air-Conditioning Engineers (ASHRAE); American Conference of Governmental Industrial Hygienists (ACGIH); and National Research Council (NRC) (see Table 2 and SI Section S1.2.5) [2,52-58]. For a sample that did not include scripted activities, this system could inform home occupants of whether, over the course of a full week, any pollutants were detected at levels that might motivate mitigation. Each gaseous pollutant was classified based on the highest 1-h block average or 8-h rolling average concentration measured during the experiment. The classification obtained from CO₂ reference monitor data was compared to the classification obtained from uncorrected CO₂ concentrations reported by the low-cost NDIR sensors in the HHBs. The classifications obtained from CO, NO₂, and O₃ reference monitor data were compared to the classifications obtained from CO, NO₂, and O₃ concentrations derived from HHB data using the best-performing Alphasense algorithms (Equations 4, 4, and 3, respectively). Each PM size fraction was classified based on the 168-h average concentrations derived from HHB filter samples.

Table 2.

Columns 3–6: Averaging periods and concentration ranges used to classify the level of each pollutant as “low,” “elevated,” or “high.” Columns 7–8: The highest 1-, 8-, or 168-h average concentration of the given pollutant. Columns 9–10: The classifications determined using the reference monitor and the Home Health Boxes. Concentrations shown in the “HHB” column were obtained using uncorrected CO₂ data; the best-performing Alphasense algorithms for CO, NO₂, and O₃ (Equation 4, Equation 4, and Equation 3, respectively); as well as PM_2.5 and PM₁₀ filter samples. For “HHB” concentrations, the first value represents the median and the values in parentheses represent the range across all eight (for CO, NO₂, and O₃) or nine (for CO₂, PM_2.5, and PM₁₀) boxes. O₃ was classified using the median HHB.

Pollutant	Units	Averaging period (h)	Range for each classification			Highest average measured		Classification
			Low	Elevated	High	Reference	HHB	Reference	HHB
CO2	ppm	8	< 1200	1200 – 5000	> 5000	2198	1597 (1494, 1745)	Elevated	Elevated
CO	ppm	8	< 2	2 – 9	> 9	6	6 (6, 6)	Elevated	Elevated
CO	ppm	1	< 6	6 – 35	> 35	10	8 (8, 9)	Elevated	Elevated
NO2	ppb	1	< 53	53 – 100	> 100	219	246 (211, 271)	High	High
O3	ppb	8	< 20	20 – 70	> 70	31	59 (2, 74)	Elevated	Elevated
PM2.5	μg m−3	168	< 12	12 – 25	> 25	NA	19 (17, 21)	NA	Elevated
PM10	μg m−3	168	< 20	20 – 50	> 50	NA	40 (35, 46)	NA	Elevated

For CO₂, CO, NO₂, and O₃, we also compared each 1-h block or 8-h rolling average concentration obtained using each HHB to the corresponding reference concentration to determine the fraction of 1-h block or 8-h rolling averages for which the two concentrations resulted in the same classification. This analysis was completed twice using: (1) uncorrected CO₂ concentrations reported by the low-cost NDIR sensors in the HHBs as well as CO, NO₂, and O3 concentrations derived from HHB data using the best-performing Alphasense algorithms (Equations 4, 4, and 3, respectively) and (2) CO₂, CO, NO₂, and O₃ concentrations calculated from HHB data using the best-performing empirical calibration models (Equations 10, 9, 8, and 9, respectively).

To investigate how occupant activities influenced indoor pollutant concentrations, we grouped activities into the following categories: door/windows open, air conditioner on, air cleaner HEPA filter on, air cleaner HEPA filter + ionizer on, scripted natural gas cooking burner use (with no food cooking), cooking (with one or more natural gas burners on), and none. Scripted burner use and cooking took precedence over all other activities. An open door and windows took precedence over the air cleaner being on. In other words, if scripted burner use occurred while the door and windows were open, that period was classified as “burner on.” If the door and windows were open while the air cleaner was on, that period was classified as “door/windows open.” Our experimental design dictated that scripted burner use and normal cooking never took place at the same time. Due to normal occupant behavior, the door and windows were never open while the air conditioner was on, and the air cleaner was never on while the air conditioner was on.

We calculated a concentration-time ratio (c: t)_ik for each combination of pollutant i and activity k:

$$(c:t{)}_{ik}=\frac{{\sum }_{j\subseteq S_k}{c}_{ij}\left(t_j-t_{j-1}\right)∕{\sum }_{j=1}^J{c}_{ij}\left(t_j-t_{j-1}\right)}{{\sum }_{j\subseteq S_k}\left(t_j-t_{j-1}\right)∕{\sum }_{j=1}^J\left(t_j-t_{j-1}\right)}$$ (12)

where c_ij was the concentration of pollutant i at time j (ppb or ppm) and (t_j – t_j–1) represents the data logging interval for pollutant i (s). In the numerator of Equation 12, c_ij(t_j – t_j–1) is summed over all times associated with activity category k (i.e., j ⊆ S_k) and then divided by c_ij(t_j – t_j–1) summed over the entire duration of the experiment (i.e., j = 1 … J). In the denominator, (t_j – t_j–1) is summed over all times associated with activity category k and then divided by (t_j – t_j–1) summed over the entire duration of the experiment [4]. A concentration-time ratio >1 indicated that activity k was associated with a disproportionately high fraction of the time-integrated concentration of pollutant i, relative to the amount of time associated with activity k [4].

We calculated concentration-time ratios using: (1) CO₂, CO, NO₂, and O₃ concentrations measured using reference monitors at 60-s intervals; (2) uncorrected CO₂ concentrations reported by the low-cost NDIR sensors in the HHBs as well as CO, NO₂, and O₃ concentrations derived from HHB data using the best-performing Alphasense algorithms (Equations 4, 4, and 3, respectively) at 30-s intervals; and (3) 1-h average CO₂, CO, NO₂, and O₃ concentrations derived from HHB data using the best-performing empirical calibration models (Equations 10, 9, 8, and 9, respectively). One-hour averages spanned each clock hour (e.g., 00:00:01 to 01:00:00 HH:MM:SS). If an activity occurred during any portion of a given hour, the full hour was associated with that activity for the purpose of calculating time-activity ratios from 1 -h average concentrations. The eight (for CO, NO₂, and O₃) or nine (for CO₂) concentration-time ratios calculated for each pollutant in steps 2 and 3 were averaged. Then, we compared concentration-time ratios calculated in step 1 to concentration-time ratios calculated in steps 2 and 3, respectively, to see if reference monitor and HHB data led to the same conclusions about the relative contributions of different activity categories to indoor air pollution.

3. RESULTS AND DISCUSSION

3.1. Pollutant Concentrations over Time

Time-series plots illustrating the range of in-kitchen pollutant concentrations measured during the 168-h experiment are shown in Figure 3. As expected, scripted use of and normal cooking with natural gas burners resulted in elevated CO₂, CO, and NO₂ concentrations. The highest 15-minute average reference CO₂ concentration (3116 ppm; Table S2 and Figure S4) was measured when the broiler burner remained on for 8.3 h to simulate an all-day cooking event on October 14^th. The highest 15-minute average reference CO concentration (14 ppm) was measured during normal cooking on the evening of October 11^th.

Figure 3.

Concentrations of CO₂, CO, NO₂, O₃, and PM_2.5 measured over the duration of the experiment. Solid lines represent 15-minute average indoor concentrations reported by reference monitors. Dashed lines and gray shaded areas represent the median and total range, respectively, of 15-minute average concentrations measured using eight (for CO, NO₂, and O₃) or nine (for CO₂ and PM_2.5) Home Health Boxes (HHBs). HHB data include uncorrected CO₂ concentrations, CO and NO₂ concentrations calculated using Equation 4, O₃ concentrations calculated using Equation 3, as well as filter-corrected PMS5003-reported PM_2.5 concentrations. Dotted lines represent 1-h average concentrations reported at Colorado Department of Public Health and Environment (CDPHE) monitoring sites located 4.3 km (CO and O₃) and 3.7 km (PM_2.5) from the home. Horizontal bars along the top of each panel indicate when doors and/or windows were opened, when a central air conditioning unit was on, when an indoor air cleaner was on, when scripted use of natural gas cooking burners (which typically involved boiling water) occurred, and when normal cooking activities occurred. Bars accompanied by a shaded backdrop of the same color represent activities noted in occupants’ original log data. The reference PM_2.5 monitor (TEOM) stopped operating reliably approximately 48 h after the experiment began.

Elevated NO₂ concentrations were the clearest indicator of natural gas cooking burner use. The median 15-minute average reference NO₂ concentration was 15 ppb, but 15-minute average concentrations above 100 ppb were measured during two normal cooking activities and 4/6 scripted burner uses. One of the scripted burner uses for which the 15-minute average reference NO₂ concentration remained below 100 ppb occurred while the door and windows were open (on the afternoon of October 12^th). The 1-h average reference NO₂ concentration was above the US EPA standard of 100 ppb for 17/168 hours (Figure S5). Two of these 17 hours were associated with normal cooking—indicating that typical use of natural gas cooking burners resulted in indoor NO₂ concentrations that exceeded outdoor air quality standards.

The highest 15-minute average indoor reference O₃ concentrations (50 to 52 ppb) were measured when the door and windows were open but no natural gas cooking burners were on (i.e., no scripted use or normal cooking). When the experiment took place, Fort Collins was in a US EPA 8-h O₃ (2015) nonattainment area and the CDPHE monitor 4.3 km from the home reported 1-h average O₃ concentrations ranging from 3 to 75 ppb [59]. On October 8^th, a peak solar radiation of 698 W m⁻² was measured at the weather station at 13:10 (Figure S6), the CDPHE monitor reported a 1-h average O₃ concentration of 75 ppb at 15:00, and a peak 15-minute average indoor reference concentration of 50 ppb was measured at 16:00. Transient decreases in the indoor O₃ concentration were observed whenever natural gas cooking burners were operated (for scripted activities or normal cooking) with the door and windows closed (Figure 3). Under these conditions, O₃ reacted with nitric oxide (NO) emitted by the natural gas cooking burners to form O₂ and NO₂.

The 168-h averaged PM_2.5 concentrations derived from all nine HHB filter samples ranged from 17.2 to 20.8 μg m⁻³ (median = 19.5 μg m⁻³; mean = 19.4 μg m⁻³; relative standard deviation [RSD] = 6.1%; Table S3). The TEOM reference monitor and the real-time PM sensors inside the HHBs detected elevated indoor PM_2.5 concentrations during normal cooking (e.g., shortly after 18:00 on October 9^th, 12^th, and 13^th; Figure 3) as well as when the door and/or windows were opened. The impact of wildfire smoke on outdoor air quality throughout Fort Collins was illustrated by peak 1-h average PM_2.5 concentrations of 174, 113, 77, and 194 μg m⁻³ reported at the CDPHE monitoring site on October 8^th, 9^th, 10^th, and 14^th, respectively. On the evening of October 8^th, the reference PM_2.5 concentration increased steadily before reaching a peak 15-minute average value of 66 μg m⁻³ at 02:15 on October 9^th. This increase presumably resulted from occupants opening doors and/or windows to ventilate or cool the home. The summer and early autumn weather in Fort Collins is characterized by hot afternoons and cool nights. The temperature measured outside of the home reached a maximum 15-minute average value of 32 °C on October 8^th and a minimum 15-minute average value of 8 °C in the early morning hours of October 9^th (Figure S6). Residents often cool their homes by opening windows in the evening—either because they lack air conditioning or want to reduce the expense of operating it. A prior study also indicated that this practice can lead to elevated indoor PM_2.5 concentrations during wildfire events [8].

3.2. Sensor Calibration

3.2.1. Gas Sensors

One-hour average SCD30-reported CO₂ concentrations had strong, positive, linear correlations with the reference CO₂ concentration (Pearson’s r ≥ 0.97); however, mean bias error (MBE) ranged from −495 to −334 ppm. After 1-h CO₂ concentrations were corrected using the linear mixed model (Equation 10 and Table S3), MBE ranged from 0 to 2 ppm (Table 3 and Figure S7).

Table 3.

Agreement between 1-h average concentrations calculated from low-cost sensor data and reported by reference monitors. For each pollutant/equation/HHB combination, the root-mean-square error (RMSE) was calculated using all 1-h concentration pairs in the 168-h experiment (for CO₂, CO, NO₂, and O₃) or the first 48 h of the experiment (for PM_2.5). The same was done for the mean absolute error (MAE), mean bias error (MBE), normalized mean absolute error (NMAE), and Pearson’s r. All values listed below represent the median and total range of each metric across N HHBs. For CO₂ and PM_2.5, rows with no equation number represent uncorrected concentrations output by the sensor. For CO, NO₂, and O₃, bold text denotes the best-performing Alphasense algorithm (Equations 1-4) and empirical calibration model (Equations 7-10).

Pollutant	Unit	Sensor	Eq.	N	RMSE (unit)			MAE (unit)			MBE (unit)			NMAE (%)			Pearson’s r (−)
					Med.	Min.	Max.	Med.	Min.	Max.	Med.	Min.	Max.	Med.	Min.	Max.	Med.	Min.	Max.
CO2	ppm	SCD30	-	9	408	349	504	398	335	495	−398	−495	−334	46	38	57	0.99	0.98	>0.99
CO2	ppm	SCD30	10	9	58	40	103	37	28	65	1	0	2	4	3	6	0.99	0.97	>0.99
CO	ppb	CO-B4	1	8	1159	512	1654	969	369	1425	795	66	1384	125	38	191	0.88	0.83	0.96
CO	ppb	CO-B4	2	7	650	542	1072	507	419	927	−140	−909	245	52	39	87	0.94	0.92	0.97
CO	ppb	CO-B4	3	8	710	505	784	543	365	600	268	−234	387	64	29	76	0.92	0.90	0.97
CO	ppb	CO-B4	4	8	645	458	779	482	339	606	230	−192	446	57	28	79	0.94	0.92	0.97
CO	ppb	CO-B4	7	8	657	405	725	496	314	563	−33	−42	−23	52	30	57	0.92	0.90	0.97
CO	ppb	CO-B4	8	8	664	405	730	499	318	568	−34	−43	−26	53	31	57	0.92	0.90	0.97
CO	ppb	CO-B4	9	8	572	381	624	445	263	494	−34	−38	−27	47	24	52	0.94	0.93	0.97
NO2	ppb	NO2-B43F	1	8	21	20	43	18	17	39	−14	−16	39	84	83	248	0.94	0.92	0.96
NO2	ppb	NO2-B43F	2	8	20	18	30	17	15	24	−12	−16	24	83	70	150	0.95	0.93	0.97
NO2	ppb	NO2-B43F	3	8	19	16	30	15	12	22	−9	−21	−3	77	55	90	0.95	0.89	0.97
NO 2	ppb	NO2-B43F	4	8	22	15	29	14	10	23	−2	−13	21	63	44	139	0.94	0.91	0.96
NO2	ppb	NO2-B43F	7	8	18	15	20	11	9	13	−1	−1	−1	41	35	50	0.92	0.89	0.95
NO 2	ppb	NO2-B43F	8	8	14	11	16	8	7	19	−1	−1	−1	26	23	33	0.96	0.93	0.97
NO2	ppb	NO2-B43F	9	8	18	15	20	11	9	13	−2	−2	−1	40	35	52	0.92	0.90	0.95
O3	ppb	OX-B431	1	8	21	18	22	19	14	20	−15	−20	0	318	98	3650	−0.17	−0.29	−0.03
O3	ppb	OX-B431	2	8	21	14	23	16	9	20	−7	−20	11	1630	121	6826	−0.15	−0.31	0.07
O 3	ppb	OX-B431	3	8	21	16	26	15	12	20	−5	−20	13	667	96	4858	0.02	−0.27	0.31
O3	ppb	OX-B431	4	8	22	15	23	20	13	21	−18	−20	−4	999	100	5545	−0.27	−0.41	0.27
O3	ppb	OX-B431	7	8	8	8	9	6	6	7	−1	−2	−1	1224	412	2031	0.56	0.48	0.62
O3	ppb	OX-B431	8	8	7	7	8	5	5	6	−2	−2	−2	335	129	744	0.69	0.62	0.75
O 3	ppb	OX-B431	9	8	7	6	7	5	5	5	−2	−2	−2	175	96	1284	0.75	0.73	0.78
PM2.5	μg m−3	PMS5003	−	9	32	17	41	25	13	33	24	12	33	69	39	93	0.96	0.96	0.97
PM2.5	μg m−3	PMS5003	11	9	6	5	8	5	4	7	1	−2	4	20	19	23	0.96	0.96	0.97
PM2.5	μg m−3	SPS30	−	5	5	4	8	4	3	6	−1	−6	0	17	14	23	0.97	0.96	0.97
PM2.5	μg m−3	SPS30	11	5	5	4	10	4	3	8	1	−3	5	16	15	26	0.97	0.96	0.97

We hypothesize that the SCD30 underestimated the reference CO₂ concentration because the atmospheric pressure (~85 kPa), and thus the ambient air density (~1.0 kg m⁻³ at 25 °C), in Fort Collins (elevation 1525 m) is lower than at sea level. Lower air density translates into fewer molecules per unit volume. The SCD30 sensor includes a pressure compensation feature that should account for deviations between the ambient pressure and atmospheric pressure at sea level if the sensor is provided with ambient pressure data from an external source [60]. During this experiment, pressure data were not provided to the SCD30 and the pressure compensation feature remained deactivated. The accuracy of SCD30-reported CO₂ concentrations could likely be improved by providing the sensor with ambient pressure data.

One-hour average CO concentrations calculated using all four Alphasense algorithms were strongly correlated with the reference CO concentration (Pearson’s r ≥ 0.83; Table 3; Figure 4). Algorithm 1 overestimated the reference CO concentration (MBE > 0 for all eight HHBs), likely because (AE_u – AE_e) was almost always negative and the correction factor n_T was always positive. Equation 2 could not be applied to one CO-B4 sensor with AE₀ = 0. The 1-h average CO concentrations calculated using Equation 4—which did not account for the auxiliary electrode output—had the smallest median RMSE and MAE values. Of the empirical calibration models, only Equation 9—which included the working electrode voltage and RH as predictors—produced smaller median RMSE and MAE values than Equation 4.

Figure 4.

Comparison of 1-h average concentrations calculated from low-cost sensor data in the Home Health Boxes and measured using reference monitors (CO: TSI QTrak 7575-X with model 982 probe; NO₂: Thermo Environmental Instruments 42C; O₃: Thermo Environmental Instruments 49C). Equations 1-4 were Alphasense algorithms. Equations 7-9 were empirical calibration models. Each HHB is represented by a different color. The dashed line is y = x.

One-hour average NO₂ concentrations calculated using all four Alphasense algorithms were strongly correlated with the reference NO₂ concentration (Pearson’s r ≥ 0.89). Of the Alphasense algorithms, Equation 4—which did not account for the auxiliary electrode output—produced the smallest median MAE, MBE, and NMAE values. When natural gas cooking burners were used, peak 15-minute and 1-h average NO₂ concentrations calculated from HHB data using Algorithm 4 were higher than the corresponding reference NO₂ concentrations (Figure 3 and Figure S5). These discrepancies might have been observed because the HHBs sampled from a location directly above the stove, whereas the reference NO₂ monitor sampled from a similar height at a location 3.2 m north of the front burners; however, the 15-minute average CO concentrations calculated from HHB data using Algorithm 4 were not consistently higher than the corresponding reference CO concentrations, even though the QTrak CO monitors were installed close to the cooktop height and 1.6 m west of the left burners. One-hour NO₂ concentrations predicted using the linear mixed model shown in Equation 8—which included the working electrode voltage and temperature as predictors—had the smallest median RMSE, MAE, MBE, and NMAE values.

One-hour average O₃ concentrations calculated using the Alphasense algorithms were weakly—and often negatively—correlated with the reference O₃ concentration. We might have had trouble obtaining reliable O₃ measurements from OX-B431 data because the sensor responds to both NO₂ and O₃. In fact, calibration data from Alphasense indicated that each OX-B431 sensor was more sensitive to NO₂ than to O₃. Average NO₂ and O₃ concentrations in the kitchen were similar (median 1-h NO₂ and O₃ = 15 and 20 ppb, respectively) but the highest NO₂ concentrations were much greater than the highest O₃ concentrations (maximum 1-h NO₂ and O₃ = 219 and 46 ppb, respectively). As a result, much of the OX-B431 sensor response was likely due to NO₂ rather than O₃; however, the negative MBE values shown in Table 3 suggest that the Alphasense algorithms over-corrected for the NO₂ response. It’s also possible that some O₃ was lost to the surface of the gas sensor housing inlet and, as a result, the OX-B431 sensors were less sensitive to O₃ when installed in the HHBs than when calibrated by Alphasense.

One-hour average O₃ concentrations predicted using empirical calibration models were moderately- to strongly-correlated with reference O₃ concentrations (Pearson’s r = 0.48 to 0.78). The linear mixed model shown in Equation 9—which included the working electrode voltage and RH as predictors—produced the strongest correlation as well as the smallest median NMAE.

Overall, we hypothesize that the following results related to gas sensor calibration might be generalizable to other studies:

1. Users of NDIR CO₂ sensors should account for differences between local atmospheric pressure and atmospheric pressure at sea level.

2. Equation 2 is not an ideal calibration algorithm for Alphasense electrochemical sensors because it cannot be used if AE₀ = 0 (i.e., if the sensor auxiliary electrode output is zero when measuring zero air, which Alphasense calibration data indicated was the case for one CO-B4 sensor tested in this study).

3. We successfully predicted gas concentrations without using electrochemical sensor auxiliary electrode output. This result might have been due to our use of the sensors in an indoor environment with relatively constant temperature (20 to 27 °C) and low RH (< 50%); however, Cross et al. reached a similar conclusion after using Alphasense CO-B4, NO2-B43F, and OX-B421 sensors in an outdoor study that spanned multiple seasons [28].

4. Even in an indoor setting with relatively constant temperature and low RH, electrochemical gas sensor performance might be influenced by environmental conditions, as suggested by the fact that our best-performing empirical calibration model for each sensor included either temperature or RH as a predictor.

We compared performance metrics for the low-cost gas sensors in the HHBs to results from several studies in SI Section S2.2 [27-29,31,34,61]. The RMSE for 1-h average NO₂ concentrations calculated from Alphasense NO2-B43F sensor data using Equation 8 (median = 14 ppb) was higher than the RMSE reported by Chatzidiakou et al. [34] for 1-minute indoor NO₂ concentrations obtained from Alphasense NO2-A43F sensors calibrated using a linear model (3 ppb). A key difference is that we measured a maximum 1-minute NO₂ concentration (250 ppb) that was 6× the maximum 1-minute NO₂ concentration (40 ppb) measured by Chatzidiakou et al. The NMAEs for 1-h average CO₂ (median = 4%, range = 3% to 6%) and NO₂ (median = 26%, range = 23% to 33%) concentrations calculated using Equations 10 and 8, respectively, were similar to relative errors reported by Zimmerman et al. [29] for 15-minute average outdoor CO₂ (2%) and NO₂ (29%) concentrations obtained from low-cost NDIR and NO2-B43F sensors, respectively, that had been calibrated using random forest models. We attributed the favorable performance of the linear mixed models we used to predict CO₂ and NO₂ concentrations, compared to the more sophisticated machine learning models developed by Zimmerman et al., to several factors. We suspect that the strength of each sensors’ response to the pollutant of interest was higher and the impact of temperature and RH on sensor outputs was mitigated during our experiment because (1) normal and scripted use of natural gas cooking burners resulted in high CO₂ and NO₂ concentrations and (2) we operated the sensors in an indoor environment where temperature and RH spanned relatively narrow ranges. We also averaged data over a 4× longer period (1 h) than Zimmerman et al. (15 minutes). Results from multiple prior studies indicated that hybrid calibration models for low-cost NDIR and electrochemical gas sensors—in which a machine learning model predicted concentrations below a certain threshold (e.g., 90% of the maximum in the training dataset) and a linear model predicted higher concentrations—predicted reference gas concentrations more accurately than linear calibration models [30,31].

3.2.2. PM Sensors

The 168-h average uncorrected PM_2.5 concentrations reported by the nine PMS5003 sensors in the HHBs ranged from 26.0 to 38.5 μg m⁻³ (median = 33.2 μg m⁻³; mean = 33.4 μg m⁻³; RSD = 10.5%). The PM_2.5 gravimetric correction factor for the PMS5003 ranged from 0.54 to 0.66 (median = 0.59; mean = 0.58). Correcting PMS5003-reported PM_2.5 concentrations to the filter samples reduced the median MBE associated with 1-h average concentrations, relative to TEOM data recorded during the first 48-h of the experiment, from 24 to 1 μg m⁻³ (Table 3 and Figure S9).

Given that the RH in the kitchen remained <40% for >99% of the experimental duration, and all PM_2.5 filter samples were pre- and post-weighed in an environment with RH between 30% and 40% [43], we do not suspect that the PMS5003 sensors in the HHBs overestimated in-kitchen PM_2.5 concentrations due to hygroscopic particle growth.

The (a) minimum (0.54) and (b) maximum (0.66) gravimetric correction factors calculated across the nine PMS5003 sensors in the HHBs during the full 168-h experiment were approximately equal to (a) the 75^th percentile correction factor reported for outdoor PMS5003 sensors that measured smoke from the 2018 Camp Fire in California, USA (0.53) [62] and (b) the gravimetric correction factor calculated for the PMS5003 sensor in the outdoor ASPEN Box during 33 h at the end of the experiment (0.67; due to a firmware error, the ASPEN PM_2.5 filter sample did not run for the full 168 h). Additionally, the RMSE (3.9 μg m⁻³), MAE (3.1 μg m⁻³), MBE (1.5 μg m⁻³), and NMAE (10%) values associated with 24-h average PM_2.5 concentrations obtained via gravimetric correction of the HHB PMS5003 data were only slightly lower than the RMSE (4.5 μg m⁻³), MAE (3.5 μg m⁻³), MBE (2.6 μg m⁻³), and NMAE (11%) values associated with 24-h average PM_2.5 concentrations obtained by applying the U.S.-wide correction equation developed by Barkjohn et al. [51] for outdoor PurpleAir monitors to the HHB PMS5003 data. Overall, PM_2.5 corrections for the PMS5003 sensors in the HHBs were similar to PM_2.5 corrections for PMS5003 sensors installed outdoors during this and prior experiments; however, outdoor correction factors might not be applicable to all indoor PM_2.5 measurements. We suspect that much of the indoor PM_2.5 sampled during this experiment was wildfire smoke from outdoors. In the absence of a large, outdoor source of PM_2.5 pollution, indoor PM_2.5 pollution could be dominated by indoor sources (e.g., emissions from cooking).

The 168-h average uncorrected PM_2.5 concentrations reported by the nine SPS30 sensors varied substantially (range = 3.9 to 19.2 μg m⁻³). Four SPS30 sensors reported 168-h average uncorrected PM_2.5 concentrations between 3.9 and 11.4 μg m⁻³, while the other five reported 168-h averages between 16.4 and 19.2 μg m⁻³. During prior laboratory testing of the SPS30, we observed high precision between eight replicate sensors [18]; as a result, we suspected that the SPS30 sensors in four of the nine HHBs were not functioning properly—potentially due to the way the sensors were installed in or powered by the HHBs. PM_2.5 gravimetric correction factors for the five properly functioning SPS30 sensors ranged from 0.95 to 1.27 (median = 1.09). Correcting PM_2.5 concentrations reported by these five SPS30 sensors to the filter samples had little effect on the RMSE, MAE, MBE, or NMAE (Table 3 and Figure S9).

3.3. Classification of Pollutant Concentrations

Pollutant concentrations derived from HHB gas sensor data—using uncorrected CO₂ concentrations, CO concentrations calculated using Equation 4, NO₂ concentrations calculated using Equation 4, and O₃ concentrations calculated using Equation 3—resulted in the same classifications as concentrations derived from reference instrument data (Table 2). CO₂ concentrations were “elevated”, CO concentrations were “elevated”, and NO₂ concentrations were “high”. The indoor O₃ concentration was classified as “elevated” based on the reference data. The fact that the highest 8-hour O₃ concentration from the median HHB also fell within the “elevated” range might have been coincidental, given that 1-h O₃ concentrations calculated using Equation 3 were weakly correlated with 1-h reference concentrations (Table 3).

The results shown in Table 2 do not guarantee that classifications derived from HHB data and reference monitor data would agree for any given 168-h sample. In SI Section S2.3, we discuss how this agreement might be affected by the range of concentrations measured during a sample.

3.4. Concentration-Time Ratios

Both plots shown in Figure 5 feature concentration-time ratios obtained from reference data on the x-axis. The plot in the left panel of features concentration-time ratios obtained from HHB data consisting of uncorrected CO₂ concentrations as well as CO, NO₂, and O₃ concentrations calculated using the best-performing Alphasense algorithms (Equations 4, 4, and 3, respectively) on the y-axis; the plot in the right panel features concentration-time ratios obtained from HHB data consisting of CO₂, CO, NO₂, and O₃ concentrations calculated using the best-performing empirical calibration models (Equations 10, 9, 8, and 9, respectively) on the y-axis. Data along the diagonal dashed line indicated that equivalent ratios were obtained from reference monitors and HHBs. For points in the upper-right quadrant formed by the two dotted lines, reference and HHB data both indicated that the activity was associated with a disproportionately high fraction of the 168-h integrated concentration of the pollutant. For points in the upper-left quadrant, HHB data suggested that the activity was associated with a disproportionately high fraction of the 168-h integrated concentration of the pollutant, but reference data indicated that it was not (i.e., HHB data produced a “false-positive” result). For points in the bottom-right quadrant, reference data indicated that the activity was associated with a disproportionately high fraction of the 168-h integrated concentration of the pollutant, but HHB data did not (i.e., HHB data produced a “false-negative” result). For points in the bottom-left quadrant, reference and HHB data both indicated that the average concentration of the pollutant during the activity was lower than the 168-h average concentration of the pollutant.

Figure 5.

Left: Concentration-time ratios were obtained from HHB data consisting of uncorrected CO₂ concentrations as well as CO, NO₂, and O₃ concentrations calculated using the best-performing Alphasense algorithms (Equations 4, 4, and 3, respectively). Right: Concentration-time ratios were obtained from HHB data consisting of CO₂, CO, NO₂, and O₃ concentrations calculated using the best-performing empirical calibration models (Equations 10, 9, 8, and 9, respectively). In both panels, the dashed diagonal line is y = x and dotted lines represent concentration-time ratios = 1.

Whether or not empirical calibration models were applied to HHB data, reference monitors and HHBs both indicated that (a) scripted use of gas cooking burners contributed disproportionately to time-integrated CO₂, CO, and NO₂ concentrations and (b) normal cooking with gas burners contributed disproportionately to time-integrated CO and NO₂ concentrations. When empirical calibration models were applied to HHB data, agreement between the concentration-time ratios calculated from reference and HHB data improved—especially for O₃.

4. CONCLUSIONS

One week of data collected in a home kitchen in Fort Collins, CO, USA indicated that, relative to ASHRAE, ACGIH, US EPA, and WHO air quality standards, CO₂ concentrations were elevated (maximum 8-h reference = 2198 ppm), CO concentrations were elevated (maximum 8-h and 1-h reference = 6 and 10 ppm, respectively), and NO₂ concentrations were high (maximum 1-h reference = 219 ppb) due, in part, to scripted use of natural gas cooking burners; however, when only data affected by normal occupant activities—which included cooking with natural gas burners—were considered, CO₂ concentrations remained elevated (maximum 8-h reference = 1241 ppm), CO concentrations remained elevated (maximum 8-h and 1-h reference = 2 and 9 ppm, respectively), and NO₂ concentrations remained high (maximum 1-h reference = 139 ppb). Ozone, PM_2.5, and PM₁₀ concentrations were elevated (maximum 8-h reference O₃ = 31 ppb; median 168-h PM_2.5 = 19.5 μg m⁻³, RSD = 6.1%, n = 9 HHB filter samples; median 168-h PM₁₀ = 40.0 μg m⁻³, RSD = 7.6%, n = 9 HHB filter samples) due primarily to infiltration of outdoor air pollution—which included wildfire smoke—when windows and doors were opened.

Sensor manufacturers’ calibrations of the low-cost CO₂, CO, and NO₂ sensors inside the Home Health Boxes were accurate enough to produce the same qualitative conclusions as reference monitors regarding (a) whether concentrations of these pollutants were low, somewhat elevated, or concerningly high relative to air quality guidelines and (b) which in-home activities were associated with disproportionately high fractions of the 168-h integrated concentrations of these pollutants. Our results do suggest, however, that empirical calibration (e.g., through collocation with reference monitors) would be necessary to obtain accurate quantitative data from low-cost gas sensors. Empirical linear calibration models with sensor-specific coefficients (a) reduced the error associated with CO₂, NO₂, and O₃ concentrations obtained from the HHB data and (b) improved agreement between the HHB and reference data regarding the extent to which different activities contributed to the time-integrated concentration of each pollutant.

This study had several limitations. First, the CO “reference” monitors (TSI QTrak) relied on the same measurement technique (electrochemical) as the low-cost CO sensors; the low-cost CO sensors in the HHBs might not agree as well with a regulatory-grade NDIR CO analyzer. Second, the monitors did not all sample from the same exact location within the kitchen. Data were averaged over 15-minute intervals prior to plotting results and over 1-h intervals prior to calculating sensor performance metrics, and these averaging times likely attenuated differences in instrument response timing resulting from differences in location; however, we would have preferred to install the instruments closer together. Third, we only collected data over a single week in a single American home; therefore, we did not investigate how well the calibration models tested here transfer to other homes or retain their predictive abilities over time (e.g., months following initial calibration). Poorer sensor and calibration model performance might be observed in homes with no or less consistent climate control.

Finally, our study evaluated the accuracy and precision of pollutant concentrations measured using HHBs but did not evaluate whether knowledge gained from HHB data helped occupants or homeowners take actions to reduce occupants’ exposure to air pollution. As our data demonstrate, indoor air pollutant concentrations are influenced by occupant activities, but some pollutant sources are outside of occupant control. For example, occupants cannot control outdoor air pollution from wildfires, traffic, and industrial sources—all of which can infiltrate the home. In the case of indoor pollution sources and sinks, such as natural gas cooking burners and exhaust hoods, renters might have no control over the presence or absence of these items and be limited in their ability to move. Homeowners might not be able to afford to replace a gas stove or install a new exhaust hood. Despite these limitations, HHB data might help air quality professionals and home occupants identify the largest sources of indoor air pollution in a home. This knowledge, in turn, could inform cost-effective behavioral changes, ventilation strategies, and filtration strategies to mitigate occupants’ exposure to indoor air pollution.

HIGHLIGHTS.

• We assembled a platform to sample/sense PM_2.5, PM₁₀, CO₂, CO, NO₂, and O₃ indoors

• Nine units were collocated with reference monitors in a home kitchen for one week

• We tested manufacturer and empirical linear calibrations for low-cost gas sensors

• Gas sensors correctly identified whether pollutants exceeded air quality guidelines

• Cooking with natural gas burners led to 1-h indoor NO₂ concentrations above 100 ppb