AAAR 37th Annual Conference October 14 - October 18, 2019 Oregon Convention Center Portland, Oregon, USA
Abstract View
Improving the Performance of Low-Cost Optical Particle Counters with Machine Learning: Applications for Indoor Aerosol Measurements
Satya Sundar Patra, RISHABH RAMSISARIA, Ruihang Du, Tianren Wu, Brandon E. Boor, Purdue University
Abstract Number: 828 Working Group: Air Quality Sensors: Low-cost != Low Complexity
Abstract Machine learning-based calibration techniques are emerging as a viable option to improve the field performance of low-cost optical particle counters (OPCs). Raw particle number and mass concentration data provided by low-cost OPCs is often inaccurate and subject to uncertainties. The objective of this study is to apply machine learning techniques to improve the accuracy of low-cost OPCs in measuring accumulation and coarse mode aerosols in residential buildings.
A two-month field campaign was conducted from November 2018-January 2019 at the Purdue ReNEWW (Retrofitted Net-zero Energy, Water, and Waste) House to collect training and testing data for low-cost OPCs. Three adult residents occupied the house throughout the campaign. The OPCs report raw particle counts across sixteen size fractions from 380 to 17,000 nm in optical equivalent diameter. The OPCs were integrated with Raspberry Pi running a custom Python script. Co-located training aerosol instrumentation included a scanning mobility particle sizer (SMPS: 10 to 380 nm) with a long differential mobility analyzer and an optical particle sizer (OPS: 380 to 10,000 nm). The field campaign included a one-week training period and a seven-week testing period. Throughout both periods, the OPCs sampled aerosols of indoor and outdoor origin with variable optical and morphological properties.
The machine learning-based field calibration method was implemented during the one-week training period and included two stages. First, support-vector machine and gaussian process regression models were applied to correct the size-resolved counting efficiency of the OPCs from 380 to 10,000 nm using the OPS as the reference. Then, another gaussian process regression model was used to predict the number, volume, and mass of indoor aerosols below the 380 nm detection limit of the OPCs, using the SMPS as the reference. The machine learning correction approach reduced the mean absolute percentage error for OPC-based size-integrated particle number, volume, and mass concentrations (100 to 2,500 nm) from > 80% to within 10% compared to the training instrumentation during the seven-week testing period.