AI DRIVEN EMBEDDED SYSTEM FOR REAL TIME WATER QUALITY MONITORING USING MULTI-SENSOR FUSION

Abstract

Safe water is essential for drinking, agriculture, aquaculture, and industry, yet the quality of water in lakes, reservoirs, storage tanks, and distribution pipelines can change quickly because of sewage inflow, industrial discharge, agricultural runoff, and pipeline damage. In many places water quality is still checked by collecting samples manually and testing them in a laboratory, which gives accurate results but only at long intervals. This paper presents an AI driven embedded system for real time water quality monitoring that uses multiple sensors, local data fusion, and machine learning to detect unsafe water as soon as conditions begin to change. The proposed system uses an ESP32 microcontroller connected to a set of low-cost water quality sensors, including pH, turbidity, total dissolved solids, electrical conductivity, dissolved oxygen, and water temperature probes. Readings are sampled at regular intervals, filtered, and compensated for temperature before they are combined. Instead of judging each parameter in isolation, the system fuses the readings into a single feature vector that captures how the parameters behave together, since contamination events usually change several parameters at once. A lightweight machine learning model, trained on labelled samples and converted for embedded execution, runs on the controller and classifies the water into categories such as safe, moderately polluted, and unsafe. It also computes a Water Quality Index and flags sudden changes that may indicate contamination or sensor failure. Because the classification happens on the device, the system can raise a local alarm even when the network is unavailable, and it only needs to send compact summaries to the cloud during normal operation. Data is transmitted to a cloud platform using MQTT over Wi-Fi, or over LoRa for remote water bodies where Wi-Fi is not available. A web dashboard displays live readings, trends, the water quality class, and the location of each monitoring node on a map. Automatic notifications are sent by SMS, email, and mobile application when the water becomes unsafe, allowing water supply staff, farmers, or aquaculture operators to take action before the problem spreads. Readings from all nodes are kept as a time-series history that can be compared with laboratory reports during audits. Experimental evaluation using water samples of known quality showed that the fused model classified water more accurately than threshold rules applied to individual sensors and that it produced fewer false alarms caused by single-sensor drift. The system is inexpensive, solar powered in its field version, and easy to deploy at multiple points. Future work includes adding sensors for nitrate, ammonia, and heavy metals, applying anomaly detection for early warning, and building a network of nodes that can trace the source of contamination along a river or pipeline.