Long-term deployment of potentiometric sensors for continuous monitoring applications relies on the presence of a stable reference electrode. Unfortunately, inaccurate measurements due to reference electrode drift from poisoning or electrolyte leakage is an unsolved problem. This issue necessitates ongoing sensor maintenance and calibration, thus limiting attractiveness and applicability of potentiometric sensors even when compared to costly and laborious manual sampling.
Here, we present a method for detecting reference electrode drift using an array of correction electrodes and neural net post-processing, enabling correction of potentiometric sensor measurements without a need for hands-on sensor maintenance or recalibration. Indicator and correction electrode potentials were measured simultaneously versus a common reference. This collection of potential measurements was then used to train a convolutional autoencoder, designed to separate the measured cell voltages into individual electrode potentials, thus enabling the correction of signal drift due to an unstable reference. We have found from studies with simulated and experimental data that this model is capable of correcting various types of reference electrode instability over arbitrary time scales, and that this method may be useful for correcting other types of sensor drift.