An energy efficient TinyML model for a water potability classification problem

Safe drinking water is an essential resource and a fundamental human right, but its access continues beyond billions of people, posing numerous health risks. A key obstacle in monitoring water quality is managing and analyzing extensive data. Machine learning models have become increasingly prevalent in water quality monitoring, aiding decision makers and safeguarding public health. An integrated system, which combines electronic sensors with a Machine Learning model, offers immediate feedback and can be implemented in any location. This type of system operates independently of an Internet connection and does not depend on data derived from chemical or laboratory analysis. The aim of this study is to develop an energy-efficient TinyML model to classify water potability that operates as an embedded system and relies solely on the data available through electronic sensing. When compared with a similar model functioning in the Cloud, the proposed model requires 51.2% less memory space, performs all inference tests approximately 99.95% faster, and consumes about 99.95% less energy. This increase in performance enables the classification model to run for years in devices that are very resource-constrained.