Modelling Sensors Degradation for Water Quality Monitoring: A Data-Driven Approach

Sensors fault is a critical issue for any kind of application. Abrupt sensor fault are frequent, but the most common situation shows sensors that slowly vary their sensing capabilities along time, generating the so-called Concept Drift (CD). CD handling techniques are based on the possibility to collect new supervised data and estimate the drift effect according to their time variation. The Water Quality Monitoring (WQM) field is affected by the issue since electrochemical sensors degradation is common due to corrosion agents and the difficulties in periodically reaching each node of the WQM sensor network for proper maintenance. This paper discusses a possible approach for CD handling in WQM by using Machine Learning (ML) classifiers as "fake-supervisors" for labeling incoming new data. To this aim, the classifier must be reliable inside a certain neighborhood of the original training dataset centroids. Hence, by collecting data through a commercial WQM sensor, we simulate the aging process by means of an electrochemical corrosion model. Different types of corrosion trends are considered. The performances drop of three popular ML classifiers (namely Multy-Layer Perceptron Neural Networks, K-Nearest Neighbors, and Decision Trees) are compared when both raw data and data processed with PCA transformation are evaluated. Experimental results show that Neural Networks appear more robust to data perturbation, moreover the use of a PCA pre-processing step have a considerable positive effect in the model performance degradation trend.