The Outdoor Dust Information Node ( ODIN ) – development and performance assessment of a low cost ambient dust sensor

Dust Information Node (ODIN) as well as the results of the first real-world measurements. The lab tests indicate that the Sharp dust sensor used in the ODIN presents a stable baseline response only slightly affected by ambient temperature. The field tests indicate that ODIN data can be used to estimate hourly and daily PM2.5 concentrations after appropriate temperature and baseline corrections are applied. The ODIN seems 10


Introduction
Global estimates indicate that between 70-90 % of the population is exposed to average annual PM 2.5 concentrations that exceed the World Health Organization Air Quality Guideline of 10 µg m −3 (annual average).This translates into between 2.2 and 7 million premature deaths year −1 worldwide (Brauer et al., 2012;Lim et al., 2012).
A significant source of uncertainty in these estimates is the spatial representativeness of the measurements used to generate them.The cost of installing and operating air quality monitoring stations means that a small number of measurements are used to represent extensive geographical areas (Wilson et al., 2005;Asian Development Bank, 2014).This is particularly relevant in urban areas where large gradients are expected in the concentration of pollutants especially in areas with large density of sources.These gradients can not be resolved if the spatial coverage of the measurements is inadequate (Chow et al., 2002;Holstius et al., 2014;Wilson et al., 2005).
Also, capturing accurate data to assess personal exposure to air pollution is critical when dealing with human health effects (McKone et al., 2008;Snyder et al., 2013) Determining exposure in both time and space is challenging as individual behaviour affects daily and long term exposure to air pollution (Wilson et al., 2005).Therefore, health effect studies require accurate, spatially and temporally resolved air quality data.Epidemiological studies often use ambient concentrations at a city level as a proxy for exposure.The spatial and temporal resolution of reported data sets means they are unlikely to adequately reflect individual pollution exposure (Chow et al., 2002).
Modelling can partially compensate for the lack of spatial resolution, but in regions with complex topography and meteorology or unevenly distributed and poorly characterised emission sources, model accuracy is constrained (Wilson et al., 2005).
The spatial and temporal resolution of air pollution data sets could be improved by deploying additional sensors to supplement existing monitoring networks.Where previously the cost of doing so was prohibitive, rapid advances in technology in recent years have seen increasing numbers of low cost air pollution sensors available.Individuals and non-regulatory groups have seized the opportunity and developed citizen science initiatives to monitor their local air quality (Egg, 2014;Hart and Martinez, 2006;Smith and Clark, 2013;SPECK, 2015).Whilst low-cost sensors have increased the accessibility of air pollution data to the general public, the sensors are not without limitations.Instruments used for regulatory monitoring must meet high standards of precision, accuracy, comparability and traceability (Wang and Brauer, 2014) whereas low cost sensors are often provided without calibration information and have either not been characterised under ambient conditions or tested only for a limited time leaving questions on their long term reliability (Snyder et al., 2013).This is not necessarily a problem if the purpose of the measurements is commensurate with the capabilities of the sensors.For example, in some cases qualitative information indicating an increase or decrease of pollutant levels might be sufficient.However, in order to derive quantitative information from each sensor response, some level of evaluation against more robust monitors is required.This is likely to result in site specific calibrations, given the variable nature of aerosol Figures

Back Close
Full composition.To date very few low-cost pollution sensors have been evaluated against compliance monitoring instruments (Holstius et al., 2014;Williams et al., 2014).
We have made some progress in the use of low cost sensors for the assessment of personal pollution exposure with the development of the Particles, Activity and Context Monitoring Autonomous Node (PACMAN) for indoor exposure studies (Olivares et al., 2013).In this work we describe the development of the outdoor counterpart of PAC-MAN the Outdoor Dust Information Node (ODIN), a new, low-cost, battery powered sensor package to monitor outdoor particulate concentrations.We also explore the performance of the dust sensor at the heart of ODIN and PACMAN in terms of baseline stability.Finally, we present results of the first field tests of the ODIN co-located with traditional standard instrumentation at a regulatory monitoring site in Christchurch, New Zealand.These tests indicate that the ODIN is able to capture PM 2.5 concentrations once suitable corrections are applied which make it a viable instrument to measure PM in combustion dominated areas complementing traditional measurements.

ODIN
The Outdoor Dust Information Node (ODIN) was developed leveraging on the extensive open source hardware and software communities using readily available components.In very broad terms, the ODIN is a set of sensors that are integrated by a microcontroller that logs the data.Figure 1 shows a diagram of the operation of ODIN and a view of the internal layout of the device.Power is drawn from either the battery or the solar panel depending on the sunshine.To maximize the battery life, the ODIN uses its microcontroller's sleep mode and only wakes up to take a single measurement every minute.The sensors feed information to the microcontroller which in turn saves it to the memory card.The electric design files are available from Olivares and Edwards (2014) and the detail of the components is as follow.-Dust sensor: the Sharp Optical Dust Sensor GP2Y1010AU0F was selected because of its low power consumption (SHARP, 2015).The basic measurement principle of this sensor is infra-red light scattering.An infra-red light emitting diode (IRLED) and a photodetector (PD) are arranged 90 • to each other around the sensing volume.The particles in the sensing volume scatter the light from the IRLED which is measured by the PD.This sensor requires 5-7 Vdc to operate and outputs a 0-3 Vdc signal proportional to the total dust mass measured.The operation of this sensor is as a 10 ms sampling cycle consisting of a high phase of 0.32 ms and a low phase of 9.68 ms.A single measurement is taken 0.28 ms into the high phase (SHARP, 2015).To smooth the output of this sensor, the microcontroller takes 100 samples of 10 ms each and records their average.
Note that the dust sensor does not have a defined measurement size range but that the upper and lower cut off sizes are controlled by the sensitivity of the internal amplifying circuit.It is expected that this circuit has significant inter-instrument variability as well as its response to depend on the type of aerosol sampled.
-Temperature and relative humidity: the AM2302 temperature-humidity sensor from AOSONG was selected for its size, power consumption and ease of interface.According to the manufacturers, the AM2302 is accurate to within 2 % RH and 0.5 • C and it is able to operate in a wide range of conditions (AOSONG, 2015).
-Microcontroller: Sparkfun's Arduino Pro Mini (Sparkfun, 2014) based on the AT-Mega328 microcontroller was selected primarily because of its ease of use and the programming libraries available for the Arduino system.The firmware used on this microcontroller is available in the project's GitHub repository (Olivares and Edwards, 2014).
-Memory: a 2 GB micro SD card is used to log the data.Adafruit's micro SD card adapter board (Adafruit, 2014) was used to interface with the microcontroller.The files created by the units are labelled as YYYYMMDD.TXT, e.g."20120520.TXT" corresponds to the file for the 20 May 2012.Introduction

Conclusions References
Tables Figures

Back Close
Full  Macetech, 2013) was used to keep track of time.This component is documented to have a drift smaller than one minute per year.A specific library was used to communicate with this component (Maks, 2012).
-Power: a lithium ion polymer battery pack (3.7 V) delivers the power required for the unit.A 6 Ah battery is capable of powering the system for six weeks.A solar panel was added to the design which under ideal conditions of sunshine, is capable of powering the system as well as recharging the battery for long term deployments.Unfortunately the chosen combination of solar panel and charging circuit was unable to recover the battery so in future revisions a new solution will be tested.
Figure 1 also shows the internal layout of the ODIN.As the ODIN is a passive instrument, i.e. there is no forced air sample, the dust sensor is located against the wall of the enclosure with its opening directly towards the outside of the instrument, protected by a rain shield and a coarse metal mesh.
The cost of the materials for the prototype ODIN is less than USD 300 (June 2015) and each unit takes about three hours to put together.

Deployments
Several short deployments were used to capture the data presented here.The first one was aimed at characterise the zero response of the dust sensor at the heart of the ODIN.Eight Sharp dust sensors were set up in a custom made, sealed and uninsulated enclosure and placed in a controlled environment for five days at a temperature of 22 ± 1 • C and a relative humidity of 50 ± 5 %.Following this deployment, the same eight units were placed outdoors in the same sealed container but exposed to sunshine.The sensors were subject to temperatures ranging from 6 to 26 • C.More than 4000 records (66 h) were obtained.Introduction

Conclusions References
Tables Figures

Back Close
Full To evaluate the response of the ODIN to woodsmoke and assess its performance against standard air quality instrumentation, one unit (ODIN_01) was located at Environment Canterbury's1 air quality monitoring site at Coles Place (−43.511236 • S; 172.633687 • W) between the 24 July and the 14 August 2014.The unit was attached to the meteorological mast at the same height as the inlets for the PM 10 and PM 2.5 instruments at the site (Fig. 2).The data were manually downloaded from the internal memory at the end of the test period.Data from the air quality monitoring station were obtained from Environment Canterbury.These data were provided before the normal QA process as 60 min moving average, every 10 min of, PM 10 , PM 2.5 measured by a TEOM-FDMS instrument, wind speed and direction and air temperature.The details of these measurements are described by Aberkane et al. (2010).

Data analysis
All the data analysis was done using R Team (2014) and the Openair R package (Carslaw and Ropkins, 2012).The raw data are available from Olivares (2015a) and the analysis scripts are available from Olivares (2015b).
The baseline of each instrument was obtained by averaging all the measurements taken during the temperature controlled baseline deployment described above.This baseline was then subtracted from the raw measurements to obtain the baseline corrected signal.The temperature response of the sensors was estimated using data from the variable temperature baseline deployment.A linear regression between the baseline corrected signal and the ambient temperature in the enclosure was performed to describe the effect of ambient temperature on the response of the Sharp dust sensors.
For the co-location deployment the data analysis included the following: we first removed baseline drift, then corrected for temperature effects and then found the calibration coefficients to approximate PM 2.5 .Figures

Back Close
Full More in detail, the baseline drift was estimated from the linear trend of the raw response from ODIN.This baseline was subtracted from the raw ODIN output.
Then, a temperature correction was applied by first finding the linear regression between the de-trended ODIN signal and the instrument temperature.This was done by considering only data above 10 • C. Figure 3 shows that above 10 • C the temperature effect dominates the ODIN response.The data were then corrected by subtracting the linear regression with the temperature from the de-trended ODIN output.
To obtain the correction coefficients to estimate PM 2.5 , this de-trended and temperature corrected ODIN signal was used in a linear regression with PM 2.5 as the dependent variable and the ODIN signal.To explore the stability of this correction, the linear regression was performed on the first 1/3 of the data while the error estimate (RMSE) was calculated using the whole dataset.

Dust sensor behaviour
Table 1 shows the results of the baseline test for the eight sensors tested and highlights the inter-instrument variability for the Sharp dust sensor with the range of baselines between 0.24 to 28 mV.For each sensor, the standard deviation of the baseline estimate suggests a relatively stable baseline for the sensors tested.
In our previous work with similar sensors we found a linear relationship between the baseline response of the dust sensor and ambient temperature in field deployments (Olivares et al., 2013).Our original hypothesis was that this interference by temperature was a property of the sensor so, as described above, we placed several units in a clean environment exposed to a range of temperatures.
If the effect of ambient temperature on the Sharp dust sensor readings was related to the sensor itself we would expect that there would be a significant increase in their baseline response in warmer temperatures even when placed in a clean environment.As shown in Fig. 4, ambient temperature appears to change the response of the sensors by around 0.3 mV for every 1 • C. This, however is a much weaker response to temperature than what is found when using the sensor as part of the ODIN.Table 2 shows that the temperature correction applied to the ambient data is of 4.4 mV for every 1 • C.This suggests that the impact of temperature on the response from the Sharp dust sensor is dominated by impacts on the aerosol instead of being a property of the devices.However, future tests will further explore this behaviour.

ODIN's performance evaluation
Figure 5 shows that the raw ODIN data show a significant baseline drift and they only capture a small fraction of the features observed in the PM 2.5 time series.The baseline correction showed a significant drift of almost −30 % during the deployment (Table 2).After removing the baseline drift, the temperature correction was estimated as 4.4 mV • C −1 (Table 2) which, as noted above, is significantly higher than what was obtained in clean conditions but similar to what was observed by our group in a previous work (Olivares et al., 2013).
Table 2 shows that, after applying the baseline and temperature corrections, a calibration factor of 0.6 and an offset of 19 µg m −3 is sufficient for the ODIN data to capture most of the variability of PM 2.5 .Figure 6 shows that the corrected ODIN data captured both the timing and intensity of all the high concentration events observed in the PM 2.5 data.However, for PM 2.5 concentrations below 25 µg m −3 the scatter plot shows a less defined relationship between ODIN and PM 2.5 .Figure 6 also shows that the correction calculated for the beginning of the deployment significantly overestimates the low concentrations observed towards the end of the dataset.This could be due to a residual baseline drift not accounted for by the method used but it does not seem to affect the fit for the higher concentrations.
Finally, and to evaluate the behaviour of the ODIN's inlet, we compared the error in the ODIN estimate (PM 2.5 − ODIN) with the observed wind speed and direction.Figure 7 shows that the error is uniform with wind direction but has a small negative trend Introduction

Conclusions References
Tables Figures

Back Close
Full with wind speed.This difference in higher wind speeds could be related to a different source mix than in low wind speeds which typically occur during night time (Aberkane et al., 2010).However, PM 2.5 concentrations are generally lower in high wind speeds which, as indicated before, is closer to the detection limit of the Sharp dust sensor.

Conclusions
The small, low-cost dust monitor ODIN, based on an optical dust sensor has been shown to be able to capture most of the features of the PM 2.5 time series in a woodsmoke impacted area after suitable baseline and temperature corrections are applied.
In controlled conditions, the optical dust sensor used in ODIN was shown to have a stable baseline with a small dependence with ambient temperature.However, field data indicates a significant baseline drift and temperature interference.The baseline drift of nearly 30 % in three weeks indicates that regular checks will be required if the ODIN is to be used for extended periods of time.
The fact that the temperature interference observed in the field tests was more significant than that observed in lab conditions suggests that the changes in the response of the ODIN to ambient temperature are related to the nature of the aerosol sampled and not only to the sensor themselves.This has implications for the transferability of the correction factors to other locations.
Also, the performance of the ODIN is worst for PM 2.5 concentrations below 25 µg m −3 .This is to be expected as the datasheet of the Sharp dust sensor has a response curve starting at 100 µg m −3 .This is a common characteristic of low-cost sensors (Wang and Brauer, 2014) and one should be cautious when using these sensors in lowconcentration environments as their response may reflect more their noise than a real measurement.
A simple test of the performance of ODIN against wind speed and direction found that the inlet performed as expected with no wind direction bias and only a small negative wind speed trend.This negative trend with wind speed may also be related to the Figures

Back Close
Full fact that high wind speeds relate to low concentrations, where the Sharp dust sensor perform worst, and that a different source mix may be dominant in those conditions.Nevertheless, the ODIN is shown to be a useful complement to regulatory measurements for campaigns and its calibration parameters are stable for deployments of around one month.
The next steps in understanding the response of the ODIN to urban aerosols are to explore more in detail the performance of the Sharp dust sensor to specific aerosol populations (size and composition), explore the inter-instrument variability and the drivers for the baseline drift and temperature interference found here.We also expect to explore the transferability of correction coefficients with data currently being captured in Auckland and Christchurch and it is expected to generate results by September 2015.
Future versions of the ODIN are expected to include distributed telemetry for high density deployments at a city scale as well as improved energy efficiency for long term deployments.Introduction

Conclusions References
Tables Figures

Back Close
Full  Full Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | -Clock: the DS3231 based temperature compensated RTC Chronodot v2.0 ( Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper |

Figure 1 .Figure 2 .Figure 3 .Figure 4 .Figure 5 .Figure 6 .Figure 7 .
Figure 1.Internal layout and simplified diagram of ODIN showing the main components and the flow of power (red arrows) and information (black arrows).The picture on the right shows how the components fit inside the enclosure with the dust sensor attached to the lid of the unit and exposed through a hole directly over its sensing area.See Olivares and Edwards (2014) for the full electrical schematics.

Table 1 .
Results from the dust sensor baseline experiment where eight Sharp dust sensors were placed in a clean, temperature controlled environment.