Towards widespread adoption of low cost air quality sensors - a necessity for effective calibration procedures
Апстракт
One of the most important factors for increasing usefulness and relevance of air pollution data on a personal level would certainly be an increase in its spatial resolution. Current state of affairs in air quality monitoring networks at state or local level is such that they typically provide a wealth of high resolution temporal data, but monitoring stations are on the other hand mainly located at a few strategically important places in urban area. This low spatial resolution is a big barrier towards providing personally relevant information to citizens which would then be able to answer questions such as: what is the level of air pollution on routes and places they frequent, what are the associated health risks, and finally what can they do about it and at what cost? Part of the solution to this complex puzzle may be in low-cost air quality sensors (LCS). LCS’s are an emerging technology and are now commercially available for gases, particulate matter and meteorological parameters in ...a wide variety of designs and capabilities. However, the data sets generated by devices composed of selected LCS are often of questionable data quality. There are some protocols for calibration LCS in laboratory and in the field, but such procedures are extensive and appropriate for testing performance only during sensor development, but are certainly not practically feasible for testing of each commercial LCS [1, 2]. Developing, optimizing, and refining experiments and statistical modelling techniques for LCS-AQ calibration and validation is the mandatory step on the route of obtaining reliable and meaningful data [3, 4]. This work elaborates an important part of LCS deployment – its effective calibration procedure. We will describe our current work on calibration procedures for sensors for gaseous pollutants: ozone and carbon monoxide and sensors for particulate matter. In validation campaigns described in this work we have used combination of quality checks and mid-level validation, together with several statistical modeling approaches in order to observe which sensors have desirable level of performance and to later derive calibration curves or more complex calibration models. Calibration models were based on simple linear regression (LR), multiple linear regression (MLR) and artificial neural networks (ANN). Mid-level validation of particulate matter low cost sensors was done via collocation with lab-grade instruments in laboratory office space in Institute Vinca and the results were collected for several weeks. Low cost instruments included two Sharp GP2Y1010AU0F compact optical dust sensors connected to Arduino platform (1 channel output), Alphasense CompactOPC sensor (16 channels from 0.38 to 17 μm) and Dylos DC1700 PM unit (2 channels corresponding to “large” and “small” particles). Lab grade instruments included TSI NanoScan SMPS Model 3910 and TSI Optical particle sizer 3330 (17 channels from 0.3um to 10um). Basic quality check of two Sharp sensors showed that they did not have malfunctions and are surprisingly reliable when used in combination with Arduino platform. Sharp sensors mutually correlate with correlation coefficient cc~0.98. In comparison with lab grade instruments, they correlate best with OPS first channel cc~0.75 and cc steadily declines for channels corresponding to larger particles. Dylos channels best correlate with OPS 0.5-0.721μm for “small” particles with cc~0.60, and with OPS 2.156 μm for “large” particles with cc~0.978. Best performing low cost sensor was certainly Alphasense OPC with cc over 0.90 for corresponding channels. Performed validation steps clearly show to which particulate matter size range individual sensor channels correspond, enabling one to derive meaningful calibration curves. LCS’s for gaseous pollutants were deployed within multi-sensor platform AQMesh. Platforms were collocated with Automatic Monitoring Station Stari Grad belonging to the State Network run by the Serbian Environmental Protection Agency (SEPA), in two 1-month periods in late summer and early autumn 2015. CO and O3 sensors in AQMesh platform passed low level validation (criteria was percentage of collected data), and were considered for development of calibration models. Choice of predictors for MLR and ANN models utilized both statistical reasoning and heuristics to avoid overfitting calibrated sensor with co-varying gas species. Improvements in sensor performance with sophisticated ANN models compared to LR were significant, resulting in relative residuals less than 15% for concentration of pollutant approximately ranging from 175 μg/m3 to 400 μg/m3 , while for O3 in the range from 40 μg/m3 to 120 μg/m3 .
Извор:
ENVIROCHEM 2018 : 8th Symposium Chemistry and Environmental Protection : program and the book of abstracts, 2018, 39-40Издавач:
- Belgrade : Serbian Chemical Society
Напомена:
- 8th Symposium Chemistry and Environmental Protection : May 30 - June 1, Kruševac, 2018.
Колекције
Институција/група
VinčaTY - CONF AU - Davidović, Miloš AU - Topalović, Dušan AU - Tasić, Viša AU - Jovašević-Stojanović, Milena PY - 2018 UR - https://vinar.vin.bg.ac.rs/handle/123456789/12690 AB - One of the most important factors for increasing usefulness and relevance of air pollution data on a personal level would certainly be an increase in its spatial resolution. Current state of affairs in air quality monitoring networks at state or local level is such that they typically provide a wealth of high resolution temporal data, but monitoring stations are on the other hand mainly located at a few strategically important places in urban area. This low spatial resolution is a big barrier towards providing personally relevant information to citizens which would then be able to answer questions such as: what is the level of air pollution on routes and places they frequent, what are the associated health risks, and finally what can they do about it and at what cost? Part of the solution to this complex puzzle may be in low-cost air quality sensors (LCS). LCS’s are an emerging technology and are now commercially available for gases, particulate matter and meteorological parameters in a wide variety of designs and capabilities. However, the data sets generated by devices composed of selected LCS are often of questionable data quality. There are some protocols for calibration LCS in laboratory and in the field, but such procedures are extensive and appropriate for testing performance only during sensor development, but are certainly not practically feasible for testing of each commercial LCS [1, 2]. Developing, optimizing, and refining experiments and statistical modelling techniques for LCS-AQ calibration and validation is the mandatory step on the route of obtaining reliable and meaningful data [3, 4]. This work elaborates an important part of LCS deployment – its effective calibration procedure. We will describe our current work on calibration procedures for sensors for gaseous pollutants: ozone and carbon monoxide and sensors for particulate matter. In validation campaigns described in this work we have used combination of quality checks and mid-level validation, together with several statistical modeling approaches in order to observe which sensors have desirable level of performance and to later derive calibration curves or more complex calibration models. Calibration models were based on simple linear regression (LR), multiple linear regression (MLR) and artificial neural networks (ANN). Mid-level validation of particulate matter low cost sensors was done via collocation with lab-grade instruments in laboratory office space in Institute Vinca and the results were collected for several weeks. Low cost instruments included two Sharp GP2Y1010AU0F compact optical dust sensors connected to Arduino platform (1 channel output), Alphasense CompactOPC sensor (16 channels from 0.38 to 17 μm) and Dylos DC1700 PM unit (2 channels corresponding to “large” and “small” particles). Lab grade instruments included TSI NanoScan SMPS Model 3910 and TSI Optical particle sizer 3330 (17 channels from 0.3um to 10um). Basic quality check of two Sharp sensors showed that they did not have malfunctions and are surprisingly reliable when used in combination with Arduino platform. Sharp sensors mutually correlate with correlation coefficient cc~0.98. In comparison with lab grade instruments, they correlate best with OPS first channel cc~0.75 and cc steadily declines for channels corresponding to larger particles. Dylos channels best correlate with OPS 0.5-0.721μm for “small” particles with cc~0.60, and with OPS 2.156 μm for “large” particles with cc~0.978. Best performing low cost sensor was certainly Alphasense OPC with cc over 0.90 for corresponding channels. Performed validation steps clearly show to which particulate matter size range individual sensor channels correspond, enabling one to derive meaningful calibration curves. LCS’s for gaseous pollutants were deployed within multi-sensor platform AQMesh. Platforms were collocated with Automatic Monitoring Station Stari Grad belonging to the State Network run by the Serbian Environmental Protection Agency (SEPA), in two 1-month periods in late summer and early autumn 2015. CO and O3 sensors in AQMesh platform passed low level validation (criteria was percentage of collected data), and were considered for development of calibration models. Choice of predictors for MLR and ANN models utilized both statistical reasoning and heuristics to avoid overfitting calibrated sensor with co-varying gas species. Improvements in sensor performance with sophisticated ANN models compared to LR were significant, resulting in relative residuals less than 15% for concentration of pollutant approximately ranging from 175 μg/m3 to 400 μg/m3 , while for O3 in the range from 40 μg/m3 to 120 μg/m3 . PB - Belgrade : Serbian Chemical Society C3 - ENVIROCHEM 2018 : 8th Symposium Chemistry and Environmental Protection : program and the book of abstracts T1 - Towards widespread adoption of low cost air quality sensors - a necessity for effective calibration procedures SP - 39 EP - 40 UR - https://hdl.handle.net/21.15107/rcub_vinar_12690 ER -
@conference{ author = "Davidović, Miloš and Topalović, Dušan and Tasić, Viša and Jovašević-Stojanović, Milena", year = "2018", abstract = "One of the most important factors for increasing usefulness and relevance of air pollution data on a personal level would certainly be an increase in its spatial resolution. Current state of affairs in air quality monitoring networks at state or local level is such that they typically provide a wealth of high resolution temporal data, but monitoring stations are on the other hand mainly located at a few strategically important places in urban area. This low spatial resolution is a big barrier towards providing personally relevant information to citizens which would then be able to answer questions such as: what is the level of air pollution on routes and places they frequent, what are the associated health risks, and finally what can they do about it and at what cost? Part of the solution to this complex puzzle may be in low-cost air quality sensors (LCS). LCS’s are an emerging technology and are now commercially available for gases, particulate matter and meteorological parameters in a wide variety of designs and capabilities. However, the data sets generated by devices composed of selected LCS are often of questionable data quality. There are some protocols for calibration LCS in laboratory and in the field, but such procedures are extensive and appropriate for testing performance only during sensor development, but are certainly not practically feasible for testing of each commercial LCS [1, 2]. Developing, optimizing, and refining experiments and statistical modelling techniques for LCS-AQ calibration and validation is the mandatory step on the route of obtaining reliable and meaningful data [3, 4]. This work elaborates an important part of LCS deployment – its effective calibration procedure. We will describe our current work on calibration procedures for sensors for gaseous pollutants: ozone and carbon monoxide and sensors for particulate matter. In validation campaigns described in this work we have used combination of quality checks and mid-level validation, together with several statistical modeling approaches in order to observe which sensors have desirable level of performance and to later derive calibration curves or more complex calibration models. Calibration models were based on simple linear regression (LR), multiple linear regression (MLR) and artificial neural networks (ANN). Mid-level validation of particulate matter low cost sensors was done via collocation with lab-grade instruments in laboratory office space in Institute Vinca and the results were collected for several weeks. Low cost instruments included two Sharp GP2Y1010AU0F compact optical dust sensors connected to Arduino platform (1 channel output), Alphasense CompactOPC sensor (16 channels from 0.38 to 17 μm) and Dylos DC1700 PM unit (2 channels corresponding to “large” and “small” particles). Lab grade instruments included TSI NanoScan SMPS Model 3910 and TSI Optical particle sizer 3330 (17 channels from 0.3um to 10um). Basic quality check of two Sharp sensors showed that they did not have malfunctions and are surprisingly reliable when used in combination with Arduino platform. Sharp sensors mutually correlate with correlation coefficient cc~0.98. In comparison with lab grade instruments, they correlate best with OPS first channel cc~0.75 and cc steadily declines for channels corresponding to larger particles. Dylos channels best correlate with OPS 0.5-0.721μm for “small” particles with cc~0.60, and with OPS 2.156 μm for “large” particles with cc~0.978. Best performing low cost sensor was certainly Alphasense OPC with cc over 0.90 for corresponding channels. Performed validation steps clearly show to which particulate matter size range individual sensor channels correspond, enabling one to derive meaningful calibration curves. LCS’s for gaseous pollutants were deployed within multi-sensor platform AQMesh. Platforms were collocated with Automatic Monitoring Station Stari Grad belonging to the State Network run by the Serbian Environmental Protection Agency (SEPA), in two 1-month periods in late summer and early autumn 2015. CO and O3 sensors in AQMesh platform passed low level validation (criteria was percentage of collected data), and were considered for development of calibration models. Choice of predictors for MLR and ANN models utilized both statistical reasoning and heuristics to avoid overfitting calibrated sensor with co-varying gas species. Improvements in sensor performance with sophisticated ANN models compared to LR were significant, resulting in relative residuals less than 15% for concentration of pollutant approximately ranging from 175 μg/m3 to 400 μg/m3 , while for O3 in the range from 40 μg/m3 to 120 μg/m3 .", publisher = "Belgrade : Serbian Chemical Society", journal = "ENVIROCHEM 2018 : 8th Symposium Chemistry and Environmental Protection : program and the book of abstracts", title = "Towards widespread adoption of low cost air quality sensors - a necessity for effective calibration procedures", pages = "39-40", url = "https://hdl.handle.net/21.15107/rcub_vinar_12690" }
Davidović, M., Topalović, D., Tasić, V.,& Jovašević-Stojanović, M.. (2018). Towards widespread adoption of low cost air quality sensors - a necessity for effective calibration procedures. in ENVIROCHEM 2018 : 8th Symposium Chemistry and Environmental Protection : program and the book of abstracts Belgrade : Serbian Chemical Society., 39-40. https://hdl.handle.net/21.15107/rcub_vinar_12690
Davidović M, Topalović D, Tasić V, Jovašević-Stojanović M. Towards widespread adoption of low cost air quality sensors - a necessity for effective calibration procedures. in ENVIROCHEM 2018 : 8th Symposium Chemistry and Environmental Protection : program and the book of abstracts. 2018;:39-40. https://hdl.handle.net/21.15107/rcub_vinar_12690 .
Davidović, Miloš, Topalović, Dušan, Tasić, Viša, Jovašević-Stojanović, Milena, "Towards widespread adoption of low cost air quality sensors - a necessity for effective calibration procedures" in ENVIROCHEM 2018 : 8th Symposium Chemistry and Environmental Protection : program and the book of abstracts (2018):39-40, https://hdl.handle.net/21.15107/rcub_vinar_12690 .