Calibration of Low-Cost Particulate Matter Sensors with Elastic Weight Consolidation (EWC) as an Incremental Deep Learning Method

Urban air quality is an important problem of our time. Due to their high costs and therefore low spacial density, high precision monitoring stations cannot capture the temporal and spatial dynamics in the urban atmosphere, low-cost sensors must be used to setup dense measurement grids. However, low-cost sensors are imprecise, biased and susceptible to environmental influences. While neural networks have been explored for their calibration, issues include the amount of data needed for training, requiring sensors to be co-located with reference stations for extensive periods of time. Also re-calibrating them with new data can lead to catastrophic forgetting. We propose using Elastic Weight Consolidation (EWC) as an incremental calibration method. By exploiting the Fisher-Information-Matrix it enables the network to compensate for different sources of error, both pertaining to the sensor itself, as well as caused by varying environmental conditions. Models are pre-calibrated with data of 40 h measurement on a low-cost SDS011 PM sensor and then re-calibrated on another SDS011 sensor. Our evaluation on 1.5 years of real world data shows that a model using EWC with a time period of data of 6 h for re-calibration is more precise than models without EWC, even those with longer re-calibration periods. This demonstrates that EWC is suitable for on-the-fly collaborative calibration of low-cost sensors.