P2AR.4 - Self-Calibration of Extremely Unstable Sensor Arrays
- Event
- 17th International Meeting on Chemical Sensors - IMCS 2018
2018-07-15 - 2018-07-19
Vienna, Austria - Chapter
- P-2 - Sensor Arrays
- Author(s)
- G. Magna, C. Di Natale, E. Martinelli - Dipartimento Ingegneria Elettronica, Università di Roma Tor Vergata, Rome (Italy)
- Pages
- 907 - 908
- DOI
- 10.5162/IMCS2018/P2AR.4
- ISBN
- 978-3-9816876-9-9
- Price
- free
Abstract
Classification models accomplish the recognition of unknown samples relying on a training set of data. However, since sensors are usually unstable over the long period, perpetuating the use of a same sensor array can make outdated the classification model. In these situations, if the array still maintains the capability of discriminating classes, periodical calibrations are sufficient to preserve performances. Conversely, sensor failures or malfunctioning could compromise current and further classification models; hence defective sensors have to be removed and, in case, replaced. This work investigates the implementation of a Self-Calibration algorithm aimed at detecting the occurrence of dramatic changes in the behavior of sensors. Once an anomaly is detected, the model decides either to keep or to remove the affected sensor. In the latter situation, a sensor replica is included in the array. Eventually, a model for the whole array might be trained in a totally unsupervised way and the algorithm utilizes the subset of the array still functioning to recalibrate the whole array [1]. Results show that, in terms of classification rate, the non-malfunctioning and malfunctioning scenarios are alike in case of self-calibrated models.