| dc.description.abstract | Long-term application of chemical sensor arrays for continuous monitoring is challenging as a result of sensor aging and drift. A number of techniques have been proposed to compensate for drift, but the issue remains a challenge in this domain. In addition, most drift correction approaches require periodic recalibration of the sensors, which may not be feasible for sensors deeply embedded and deployed for uninterrupted continuous monitoring. In this thesis, we propose a multi-calibration ensemble approach for compensating sensor drift. Our method characterizes drift in the sensor measurements by using past sensor measurements for which ground-truth is available and treating them as “pseudo-calibration” samples along with the recording time of those measurements. Then, we build a regression model that learns to predict the concentration of target analytes given (1) the current sensor measurements and (2) a history of these prior pseudo-calibration samples. The approach is agnostic to the particular regression method used. For this purpose, we evaluate the efficacy of the approach 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 history of prior pseudo-calibration samples, and a state-of-the-art drift correction autoencoder (DCAE) technique. We evaluated these systems on two experimental datasets from a bioprocess control application, and also characterize their performance as a function of array cross-selectivity and amount of drift in simulation.
Our proposed approach outperforms the calibration-free model and DCAE in the first experimental dataset with errors reduced by as much as 50% in some cases. The correlation between the prediction and the ground truth also improves significantly compared to the comparison methods. On the second dataset, the proposed approach show improvement in most of the cases compared to the calibration-free model. However, in comparison to the DCAE, only the neural network model shows significant improvement in some cases. In our analysis of the simulated datasets, we have found that the proposed approach shows significant robustness to the presence of drift compared to the other methods. All the three regression techniques using the proposed technique produced a drift-free performance as the amount of drift is increased in the data. As for the analysis with varying cross-selectivity in the simulated sensors, the prosed approach shows significantly lower error compared to the comparison methods. These findings indicate that the proposed technique can generate robust predictions with low error variance and can enhance the reliability of chemical sensing arrays for continuous and long-term applications. | |
