SecureIoT-FL: A Federated Learning Framework for Privacy-Preserving Real-Time Environmental Monitoring in Industrial IoT Applications

Abstract

Industrial environments face increasing challenges in achieving accurate environmental monitoring and maintaining data privacy. This paper presents the SecureIoT-FL framework, an innovative integration of federated learning and multi-sensor fusion for real-time, privacy-preserving prediction of environmental pollutants in industrial settings. A normalized Federated Averaging (N-FedAvg) as an enhanced version of the conventional FedAvg algorithm, is introduced for global model aggregation to balance the contributions from all clients with varying data volumes, thus stabilizing the learning process. This study contributes a scalable and secure solution for environmental monitoring and risk management in industrial IoT applications. The framework leverages IoT nodes equipped with Raspberry Pi 5 and sensors to detect pollutants (e.g., PM2.5, PM10, CO2, VOCs, CH₂O, CO, O₃, NO₂, H₂S, O2, CH₄, SO₂) across three distinct environments: a PCBA factory, a CNC machining factory, and a plastic injection factory. Over five months, the system collected over 516,000 data points, enabling a multi-source data fusion approach to enhance predictive accuracy. The key outcomes of the proposed work include significant improvements in the model performance over 10 biweekly training rounds, with increasing the accuracy from 70 % to 92 % and 68–89 % for LSTM and CNN models, respectively. The framework effectively predicts the pollutant concentrations over multiple timeframes (10, 30, and 60 minutes), supporting timely interventions to reduce worker exposure to harmful pollutants. The SecureIoT-FL framework employs federated learning with TLS encryption, ensuring data privacy while supporting secure local model training and weight sharing via MQTT over AWS.

Keywords

• Federated learning
• Environmental monitoring
• Industrial IoT
• Multi-sensor fusion
• Privacy-preserving data management