Federated Transformer Model for Water Contamination Detection in Distributed Monitoring Systems
Conference
Kumar, R, Soni, J, Upadhyay, H. (2026). Federated Transformer Model for Water Contamination Detection in Distributed Monitoring Systems
. 10.1109/ICAIC67076.2026.11395770
Kumar, R, Soni, J, Upadhyay, H. (2026). Federated Transformer Model for Water Contamination Detection in Distributed Monitoring Systems
. 10.1109/ICAIC67076.2026.11395770
Safety of water distribution systems is a serious societal health concern. The emergence of IoT sensors has facilitated real time monitoring, yet this creates huge time-series data that is high-dimensional and generates massive amounts of data that are often isolated across various utilities for privacy and security reasons. Such fragmentation of data prevents the creation of effective anomaly detection models. We introduce a privacy-sensitive federated learning (FL) system to water contamination detection based on Transformer-based architecture. Our solution is based on the FedAvg algorithm that allows several, distributed water utilities to jointly learn a strong global model, without ever exchanging their raw, sensitive sensor data. The self-attention module of the Transformer is particularly useful when attempting to detect intricate spatio-temporal relationships and long-range connections among various water quality metrics (e.g., pH, turbidity, chlorine), which tend to be antecedents of contamination occurrences. We show that the federated Transformer method shows high levels of detection performance, validating its potential to develop a generalized and resilient system of anomaly detection and strictly adhering to the data privacy requirements necessary to critical infrastructure.
