Machine Learning-Enhanced Calibration Algorithm for Drift Compensation in Long-Term Electrochemical Environmental Monitoring Networks

Distributed environmental monitoring using electrochemical sensors is low-cost but is subject to several issues, including drift, cross-sensitivity, and unit-to-unit variability, all of which affect long-term measurement accuracy. In this article, we present an innovative machine-learning-enhanced calibration method that combines Random Forest regression with Incremental Domain-Adversarial Networks (IDAN) and continuous online learning with recalibration based on anomaly detection. This combined approach uses different mechanisms to address the three sources of error inherent in the use of electrochemical sensors: Random Forest can capture the complex non-linear drift patterns, IDAN allows for robust adaptation of the models from one sensor or environment to another, and continuous online learning continually improves the model as new data are received. Anomaly detection enables optimization of calibration scheduling frequency. We evaluated our integrated methodology using data collected during 3–12 months of operation of electrochemical sensors measuring gases (NO2, CO, O3), ions (Cl−, NO3−, pH), metal oxides (VOCs, H2S) and amperometric sensors (DO, NH4+). Our results showed a statistically significant improvement in sensor performance. Specifically, across 10 random seeds, the mean ± standard deviation (SD) percentage reduction in root-mean-square error (RMSE) relative to traditional calibration methods was 35.2 ± 3.1%, and the mean ± SD percentage reduction in maintenance required was 24.8 ± 2.3%. Thus, our framework enables the deployment of electrochemical sensor networks for longer periods with less user intervention than existing methodologies.