Intelligent Embedded System for Real‑Time Prediction of Drinking Water Quality Using Machine Learning and Low‑Cost Hardware

Quispe-Cotacallapa G.; Lozada-Vilca R.A.; Castillo-Alarcón O.; Pampa-Vilca W.J.; Taipe-Huaman C.W. · Smart Innovation, Systems and Technologies (2026) · DOI: 10.1007/978-3-032-24262-4_34

Resumen

Monitoring drinking water quality is critical for public health; however, conventional laboratory workflows are costly, slow, and lack real-time capability, which limits timely decision-making in the field. Meanwhile, practical and low-cost tools that combine calibrated sensing with automated analytics remain scarce. This study presents the design and validation of an intelligent, portable embedded system for real-time prediction of drinking water quality. The system integrates a supervised neural network with low-cost, high-precision hardware (ESP32 + ADS1115 + Raspberry Pi 3 B+) and open-source tools (TensorFlow/Keras, TensorFlow Lite) to enable local inference (edge computing). The 16-bit ADS1115 ADC provides high-resolution digitization and a stable linear response, allowing the capture of low-level voltage variations with high sensitivity in the analog signals from the electrodes. A quantitative and experimental approach was followed, including prototype development, data acquisition and normalization (100–1400 µS/cm), training of a dense neural network (two hidden layers of three neurons, tanh activations, Adam optimizer) for 1000 epochs, and system validation. The model showed a consistent decrease in mean squared error (MSE), achieving performance with RMSE = 52 µS/cm, MAE = 35 µS/cm, and Pearson’s r = 0.985 compared to the reference conductivity meter HANNA HI98136 (relative error ≈ 2.5–3.0%). Moreover, inference latency on the Raspberry Pi 3 B+ was <50 ms, validating its real-time operation. The proposed system is portable, cost-effective, and capable of operating without cloud connectivity, offering a practical complement to conventional workflows. In addition, its modular architecture allows the analysis to be extended to other physicochemical parameters (e.g., pH, nitrates, phosphates), and its scalable field deployment contributes to decentralized and sustainable environmental monitoring. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.

Instituciones: Universidad Nacional Mayor de San Marcos; Universidad Nacional de San Agustín de Arequipa; Universidad Nacional de Juliaca