On-line drift compensation for continuous monitoring with arrays of cross-sensitive chemical sensors

Long-term application of chemical sensor arrays for continuous monitoring is challenging as a result of sensor drift. Drift correction often requires periodic recalibration, which may not be feasible for sensors deeply embedded and deployed for uninterrupted continuous monitoring. In this paper, we propose a multi-calibration ensemble approach to compensate for sensor drift in such applications. Our method uses past sensor measurements for which ground-truth is available, and treats them as “pseudo-calibration” samples. With these, it builds a regression model to predict the concentration of target analytes by combining (1) the current sensor measurements and (2) the history of prior pseudo-calibration samples. We evaluate the efficacy of the proposed model using three different regression techniques, partial least squares, extreme gradient boosting, and neural networks, and compare it against two baselines: regression models that do not use the pseudo-calibration samples, and a state-of-the-art drift-correction technique. We evaluated these models on an experimental dataset from a bioprocess control application, and characterize them as a function of cross-sensitivity in the sensor array and amount of drift through computer simulations. The proposed approach outperforms both baselines on the experimental dataset, and under all simulation conditions, achieving significantly lower normalized root mean square errors in the prediction of target variables. These results hold for the three regression models used, which indicates that the proposed approach is agnostic to the underlying regression model. • To compensate for sensor drift, chemical sensor arrays often need to be recalibrated. • Calibration with a reference analyte is infeasible if the sensors are deeply embedded in the system. • We propose a drift-compensation technique that uses past sensor measurements as pseudo-calibration samples. • Using experimental and synthetic data, the model outperforms a state-of-the-art drift compensation technique.