Machine Learning Models for Prediction of Metal Ion Concentrations in Drinking Water

We present an investigation of several machine learning (ML) models developed to predict the copper (Cu) and lead (Pb) ion concentrations in drinking water. The system where this prediction is employed is based on a microwave block loop gap resonator (BLGR), which surrounds a glass tube with drinking water passing through it. The resonator is coupled to a vector network analyzer (VNA), which collects reflection coefficient (S11) measurements over a 100 MHz - 6 GHz frequency range. It is these S11 measurements, in raw format or compressed using various signal processing techniques, that are used as input into the ML models. Our investigation looks at new convolutional neural networks (CNN) and deep neural networks (DNN) models because such models can easily be deployed on IoT microcontroller devices using tinyML technologies. Extensive simulations using real data demonstrate that DNN models that use as input features essential spectral information created from S11 traces provide performance comparable to that of CNN models but at much shorter training times and significantly smaller model sizes.