Correlation matters
As always in metrology, the estimated value must be accompanied by a statement on the uncertainty of the estimate. To be able to calculate the uncertainty accurately, it is of critical importance to consider the correlation between different sensors.
Sensor networks monitoring air quality may contain tens to hundreds of sensors.
In the FunSNM project, we have started to study the correlation structure of measurement errors of sensors (‘sensor errors’) with the help of the dataset, which contains the measurement data of 52 sensor units, which were co-located with low-uncertainty reference instruments.
One network, deployed in Helsinki, contains 25 sensors, while a network deployed in the region around Paris area, consist of 600 sensor units, out of which 100 are deployed at fixed locations and 500 on postal service vehicles. These sensor units typically measure the concentrations of pollutants like nitrogen dioxide (NO2) or particulate matter (PM) in ambient air.
The quantity of interest, i.e., the ‘measurand’ in metrological terminology, can be an interpolated value of a concentration at a non-measured location and/or at a non-measured point in time. Or the measurand can be an aggregated value, e.g., the average concentration in a certain area over a certain period.
A Gaussian process model is used to model the correlation structure of the sensor errors, and the effect of including or not including the observed correlation in the sensor errors is assessed. The results of this study will be presented at the IMEKO World Congress 2024, which will take place from August 26 till August 29 in Hamburg, Germany.