Calibration update strategies for an array of potentiometric chemical sensors

•Calibration update methods for potentiometric sensor array compared.•Detection of copper, lead and cadmium in digested wine used as a case study.•Model expansion methods perform better than signal standardization.•Weighting, Tikhonov regularization and Joint Y-PLS produce similar results. One of the obstacles to the practical use of the multisensor systems – electronic tongues is a drift, i.e. gradual change of the sensor characteristics occurring in the process of their exploitation. Sensor drift leads to the deterioration of the performance of the calibration model that was calculated before changes in the sensor responses occurred. Two approaches can be employed to deal with this phenomenon. One consists of regular re-calibration of the sensor array, which is effective but time and labor consuming. Alternatively, multivariate statistical methods can be applied to the calibration model update using small subset of the standard samples. While significant efforts have been directed to the development of the calibration transfer and calibration update techniques, they were mostly applied to the near infrared spectroscopic instruments. Very few works addressed this issue for the potentiometric sensor arrays. In the present study, applicability of calibration update methods including slope and bias correction, direct standardization using PLS2 regression and auto-associative neural network, calibration update by weighting, Tikhonov regularization and Joint Y-PLS regression to the sensor array data was evaluated. Calibration models for copper and lead quantification in mixed solutions of transition metals using an array of 7 potentiometric sensors was used as an example. Of all studied methods, weighting and JYPLS regression were shown to be the most effective permitting to reduce prediction error for new data to the level of cross-validation error for the initial data.